Organization with Incomplete Information: Essays in Economic Analysis: A Tribute to Roy Radner

Organizations with Incomplete Information ESSAYS IN ECONOMIC ANALYSIS

A Tribute to Roy Radner There have been systematic attempts over the last twenty-five years to explore the implications of decision making with incomplete information and to model an "economic man" as an information processing organism. These efforts are associated with the work of Roy Radner, who joins other analysts in this collection to offer accessible overviews of the existing literature on topics such as Walrasian equilibrium with incomplete markets, rational expectations equilibrium, learning, Markovian games, dynamic game-theoretic models of organization, and experimental work on mechanism selection. Some essays also take up relatively new themes related to bounded rationality, complexity of decisions, and economic survival. The collection overall introduces models that add to the toolbox of economists, expand the boundaries of economic analysis, and enrich our understanding of the inefficiencies and complexities of organizational design in the presence of uncertainty. Mukul Majumdar is H. T. Warshow and Robert Irving Warshow Professor of Economics at Cornell University, and Co-Editor of Economic Theory. He has also taught at Stanford University, the University of Wisconsin, Madison, and the London School of Economics. Professor Majumdar has been a Guggenheim Fellow, a Ford Rotating Research Professor at the University of California, Berkeley, an Erskine Fellow at the University of Canterbury, an Oskar Morgenstern Visiting Professor at New York University, and a lecturer at the College de France. Elected Fellow of the Econometric Society in 1976 and the author of many research articles, he is well known for his research on dynamic economic models.

Organizations with Incomplete Information ESSAYS IN ECONOMIC ANALYSIS

A Tribute to Roy Radner Edited by

Mukul Majumdar

CAMBRIDGE UNIVERSITY PRESS

CAMBRIDGE UNIVERSITY PRESS Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo, Delhi Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www. Cambridge. org Information on this title: www.cambridge.org/9780521553001 © Mukul Majumdar 1998 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 1998 This digitally printed version 2008 A catalogue record for this publication is available from the British Library ISBN 978-0-521-55300-1 hardback ISBN 978-0-521-08466-6 paperback

Contents

List of Contributors

page vii

Preface

ix

Introduction: Searching for Paradigms Mukul Majumdar

1

Equilibrium with Incomplete Markets in a Sequence Economy Wayne Shafer

20

The Existence of Rational Expectations Equilibrium: A Retrospective Beth Allen and James S. Jordan

42

Rational Expectations and Rational Learning Lawrence E. Blume and David Easley Dynamic Games in Organization Theory Roy Radner

61 110

The Equilibrium Existence Problem in General Markovian Games Prajit K. Dutta and Rangarajan K. Sundaram

159

A Practical Person's Guide to Mechanism Selection: Some Lessons from Experimental Economics Andrew Schotter

208

Organizations with an Endogenous Number of Information Processing Agents Timothy Van Zandt

239

A Modular Network Model of Bounded Rationality Kenneth R. Mount and Stanley Reiter

306

Jndex

341

Contributors

Beth Allen, University of Minnesota and Federal Reserve Bank of Minneapolis Lawrence E. Blume, Cornell University Prajit K. Dutta, Columbia University David Easley, Cornell University James S. Jordan, University of Minnesota Kenneth R. Mount, Northwestern University Roy Radner, New York University Stanley Reiter, Northwestern University Andrew Schotter, New York University Wayne Shafer, University of Illinois Rangarajan K. Sundaram, New York University Timothy Van Zandt, Princeton University

Preface

This collection of essays grew out of a conference that I organized at Cornell University in June 1992 to honor Professor Roy Radner on his sixty-fifth birthday: it is a pleasure to note that some of the distinguished alumni of Cornell (Professors Beth Allen, Prajit Dutta, James Jordan, Andrew Schotter, Raghu Sundaram) were able to contribute to this volume. Generous financial support from the Center for Applied Mathematics, the Center for Analytic Economics, and the Department of Economics is gratefully acknowledged. In this context, I am particularly grateful to David Easley, Don Randel, John Guckenheimer, Karl Shell, and Tapan Mitra for their encouragement. The success of the conference owed much to the organizational skills of Ann Stiles. I would like to thank Professors C. D. Aliprantis, A. Banerjee, J. Benhabib, J. Boyd, M. Datta, M. El-Gamal, R. Ericson, L. Fernandez, J. K. Ghosh, T. Groves, L. Hurwicz, M. Ali Khan, P. Linhart, E. Malinvaud, E. Maskin, C. B. McGuire, L. McKenzie, M. Meurer, L. Miller, K. Nishimura, Y. Nyarko, A. Radunskaya, R. Rosenthal, M. Rothschild, A. Rubinstein, G. Sorger, S. Williams, K. Yun, and E. Zajac for their help and advice at various stages of the project. This volume does not adequately reflect Roy Radner's impressive range of interests and contributions, and it is not meant to be a Festschrift. Productive as ever, Roy is keenly exploring new areas and searching for new paradigms. But time and again, he has returned to topics involving decisions and organizations with incomplete information, and we feel that a set of essays on these themes is a fitting tribute to his creativity. Ithaca, New York

IX

Introduction: Searching for Paradigms Mukul Majumdar

1

Introduction

Problems of decision making with incomplete information about all the "payoff relevant" variables have been posed and studied in depth from several angles. At one extreme are models that focus on an individual agent (e.g., the literature building on the "Savage paradigm"1). At the other extreme - in the Walrasian tradition - are models that highlight the role of prices in coordinating independent decisions made by many agents. In between, there are models - in the Cournot tradition - with a small number of interacting agents, often with divergent interests. Problems of characterizing optimal decisions of individual agents, mutual consistency of such decisions in the larger context of an organization or an entire economy, and the possibility of achieving efficiency through appropriate design of organizations have been among the themes of enduring interest. Some of the essays in this collection review recent developments on several fronts while others stress the need for new directions. 2

Equilibrium

We begin with a review of a substantial literature on economic equilibrium under uncertainty. Chapters 1 and 2 are very much in the Walrasian tradition in the sense that the decision makers participate in markets and accept the terms of transactions as given. It has been pointed out that Walrasian equilibrium theory embodies a number of ideas. Such an equilibrium portrays a balance of the forces of supply and demand I am particularly grateful to Professor Prajit Dutta for his detailed advice and criticism on an earlier version. Thanks are due to Professors Kaushik Basu, Manjira Datta, David Easley, Uri Possen, Roy Radner, and Jennifer Wissink for their comments. 1 See the recent critical assessment of Radner (1997, pp. 326-8). 1

2

Mukul Majumdar

simultaneously in all the markets. Thus, the question of existence of an equilibrium was often reduced to the study of conditions under which a certain set of equations had a solution. And John Hicks in his Value and Capital (1939, p. 60), after an exposition of the existence question by counting the number of equations and unknowns, felt that it was a great achievement of Walras "to have shown, even so schematically, the mechanism of the interrelation of markets." Developments in the fifties included a re-examination of the existence question with a striking allegiance to mathematical rigor, often in an axiomatic framework. There was also a reinterpretation of the problem (perhaps made possible by the availability of new tools) as one of proving that a number of decentralized decisions (derived as solutions to independent optimization exercises) can be carried out simultaneously.2 When one dispensed with the demand and supply functions as basic ingredients of the models of individual choice, mutual consistency of independent plans became a key element of the equilibrium concept. Hicks had earlier noted a consistency requirement as follows (1939, p. 58, footnote 1): A market is in equilibrium, statically considered, if every person is acting in such a way as to reach his most preferred position, subject to the opportunities open to him. This implies that the actions of different persons trading must be consistent. This definition of equilibrium remains meaningful even when one rejects the "price taking" behavior of Walrasian agents, and is consistent with other equilibrium concepts explored in several chapters and discussed in more detail in the later sections of this chapter. The connection between a Walrasian equilibrium and Pareto optimal allocation of resources had engaged the attention of economic theorists for quite some time, and a calculus treatment of the problem had been given by Lange (1942) and others; see the review in Samuelson (1947, Chap. 8). But, with arguments that led to some gains in generality, elegant versions of the two fundamental theorems of the new welfare economics were obtained in the fifties, and these were viewed by some as a "specification of conditions under which a decentralization of economic decisions through a price system is compatible with efficient utilization of resources" (Koopmans 1957). These welfare propositions have figured prominently in policy debates on the role of markets and government intervention, even in contexts that are far removed from the models in which they were rigorously established. They also stimulated a program of research on the more general question of attaining efficiency or Pareto 2

This point of view was cogently offered by Koopmans (1957).

Introduction: Searching for Paradigms

3

optimality through any informationally decentralized system of decision making that takes account of dispersal of information and limits on communication and information processing in an economy (and may "work" in environments where markets fail; see Hurwicz 1986 for a formal review of the literature). In the chapters by Van Zandt, Schotter, and Radner, the reader will be exposed to various facets of the problems of organizational design and possible conflicts between decentralization and attainment of efficiency.

3

Sterility?

It is not difficult to find, even in the writings of acknowledged masters of microeconomics, a sense of unease with the Walrasian theory. One does not have to go back beyond Value and Capital (1939, pp. 60-1), where it was pointed out that "many economists (perhaps most, even of those who have studied Walras seriously)" would say that Walras does give one a picture of the whole system; but it is a very distant picture, and hardly amounts to more than an assurance that things will work themselves out somehow, though it is not very clear how they will work themselves out. Other economists are theoretically less ambitious, but they at least give us results which are applicable to actual problems. Now the reason for the sterility of the Walrasian system is largely, I believe, that he did not go on to work out the laws of change for his system of General Equilibrium. A careful reading of Hicks's comments takes us out of the immediate bounds of the static Walrasian theory and prompts us to set some general criteria (that have figured prominently in many a methodological debate!) for evaluating formal economic models. First, does the model provide explanations of economic relationships and/or insights into the functioning of specific institutions? Secondly, does the model enable us to predict economic events (and prepare us better to meet contingencies)? Finally, does the model provide advice on improving the performance of an individual, an organization, or perhaps an economy? My subsequent remarks are primarily motivated by such criteria. A significant volume of research in the last thirty years has thrown further light on the points raised by Hicks regarding the scope of Walrasian analysis. Early efforts to identify conditions under which a Walrasian equilibrium is unique led to the recognition that the property holds under strong assumptions, and,... economies with multiple equilibria must be allowed for. Such economies still seem to provide a satisfactory explanation of equilibrium as well as a satisfactory foundation for the study of stability provided

4

Mukul Majumdar

that all the equilibria of the economy are locally unique. But if the set of equilibria is compact (a common situation), local uniqueness is equivalent to finiteness. One is thus led to investigate conditions under which an economy has a finite set of equilibria. (Debreu 1970)

One of Debreu's influential contributions was to point out that such flniteness holds generically over an appropriate set of models,3 and that the set of equilibrium price vectors varies continuously with the characteristics of the economy. Although no general result involving a bound on the number of equilibria was obtained, this continuity property is obviously important in the context of predicting the equilibria from the characteristics of an economy: "otherwise, the slightest error of observation on the data of the economy might lead to an entirely different set of predicted equilibria. This consideration, which is common in the study of physical systems, applies with even greater force to the study of social systems" (Debreu 1974, p. 71). This formal way of looking at the "typical" case has no doubt enriched economic theory (see e.g. Chapter 2). It is sometimes argued that multiplicity of equilibria is not necessarily bad since it holds out the possibility that there are many possible outcomes consistent with a general scenario (and enables us to catch glimpses of many worlds in a given universe!). The point is well taken and perhaps comforting. But, in the absence of some understanding of how things will work themselves out and how a particular equilibrium is likely to emerge (depending on the historical evolution) or how we can drive an economy to a "better" equilibrium, I am still left with a sense of sterility. In the Walrasian setting, the difficulties of making progress in Samuelsonian comparative statics4 have long been clear to the experts, but the limited prospects for qualitative economics was subsequently made even clearer by a class of results obtained by Sonnenschein, Mantel, Debreu, Mas-Colell, and others. We now know that any (nonInformally, finiteness was shown to be "typical": in Debreu's formulation (1970), the "exceptional" class of economies is identified with a "closed set of Lebesgue measure zero" in a finite-dimensional Euclidean space. "The usefulness of our theory emerges from the fact that by our analysis we are often able to determine the nature of changes in our unknown variables resulting from a designated change in one or more parameters. Our theory is meaningless unless it does imply some restrictions upon empirically observable quantities by which it could conceivably be refuted. This in brief is the method of comparative statics, meaning by this the investigation of changes in a system from one position of equilibrium to another without regard to the transitional process involved in adjustment" (Samuelson 1947, pp. 7-8). The difficulties in deriving comparative static results are summarized in Arrow and Hahn (1971).

Introduction: Searching for Paradigms

5

empty) compact (sub)set of the (relative interior of the) price simplex can be viewed as the set of Walrasian equilibria of some economy in which agents have continuous strictly convex and monotone preferences; see Shafer and Sonnenschein (1982, Sec. 4) for a more formal and precise discussion of such results. The challenge posed by Hicks and Samuelson on predicting the movement of variables in response to changes in parameters can be accepted only through less ambitious models with "more concrete" structures on preferences and technologies or through numerical specification of the relationships, or through drastic simplifications on the number of commodities and/or agents. Not surprisingly, all these routes have been followed in applications of Walrasian equilibrium theory, and it can be hoped that - with the dramatic advances of our computational ability - comparative statics and dynamics will be carried out through a combination of theory and computer experiments. 4

Time

Some of the themes that appear repeatedly in this volume - learning and expectation formation, for example - are essentially linked to the passage of time. Here, I would like to recall that an interesting aspect of Debreu's exposition (1959) was his emphasis on a unified treatment of time and location by defining a commodity appropriately (as long as markets are complete and the behavior assumptions hold). But how significant is such an interpretation? I am more inclined to argue that a model with "once-and-for-all" decision making (which makes it formally equivalent to a static counterpart) is inadequate to yield genuinely dynamic analysis involving intertemporal allocation and welfare.5 In order to make significant progress on the road toward a combination of welfare theory over time and the theory of capital, "it is necessary to begin from the first principles with due attention to the peculiarities of capital" (Dorfman, Samuelson, and Solow 1958, p. 416). The transition from (essentially) static to dynamic models has been attempted in several ways, with some startling conclusions, even without a recognition One can ask a similar question about the role of the very general Walrasian equilibrium model with regard to international trade. It has been argued that "if there is any one thing that could legitimately be demanded of a theory of international trade, it is that it should be capable of explaining observed patterns andflowsof trade among countries" (Chipman 1987). The Walrasian model in its general form is not particularly illuminating in this direction. In order to make headway, economists have turned to the development of systems with special structures, addressing specific issues.

6

Mukul Majumdar

of uncertainty. One can look at a sequence of momentary or temporary equilibria, where each equilibrium is defined by some consistency conditions and behavioral rules. From Malinvaud (1953) and Samuelson (1958) we learn that even in the classical convex environment, a (neverending) sequence of competitive equilibria need not be efficient in the long run. Furthermore, the "verification scenario" (Hurwicz 1986) describing the decentralized allocation in a static Walrasian economy runs into obvious difficulties, even if the correct equilibrium messages are proposed (by the mechanism designer), when no action is taken until all the verifications of intertemporal equilibrium conditions over the entire life span of the economy are completed (see Bala, Mitra, and Majumdar 1991 for details). And going beyond prices, when the agents are not clairvoyant, it may not be possible to design any intertemporally decentralized resource allocation mechanism that assures long-run Pareto efficiency.6 As a digression, I would like to recall that there was a prominent theme involving economic dynamics that came out of the Walrasian tdtonnement through Samuelson. Although the conceptual difficulties of such a framework are well known,7 the particularly appealing feature was the attempt to present an explicitly dynamic process that captured the working of markets and to look at an equilibrium as a rest point of such a process (although little was known about the efficiency or decentralizability of an economy out of equilibrium!). There was a resurgence of interest in the late fifties in such dynamics, but the mainstream of research in Walrasian economics seemed to veer away from this direction after Scarfs example of instability (I960).8 However, it seems to me that we will not meet the concerns of Hicks about understanding the working of an economy - how it evolves and how things work out by looking exclusively at a chain of sophisticated static equilibrium concepts. "How can Pareto efficiency be achieved by a decentralized price system which operates through time with no terminal date? After forty years of research, the problems raised by this question are not exhausted.... We are facing one of those cases in which scientists have to feel uncomfortable because the validity of some central propositions in their disciplines remains exposed to doubt with respect to some not fully negligible circumstances" (Malinvaud 1992). For an elaboration of these issues, see the collection of articles edited by Majumdar (1992). See Nikaido (1968), Arrow and Hahn (1971), and Hahn (1982) for analytical results and a discussion of the difficulties in interpreting the process. More recently, Saari (1985) and Bala and Majumdar (1992) have explored the possibility of chaotic trajectories even in a two-commodity model. See, however, Uzawa (1961b, 1962b), Hurwicz, Radner, and Reiter (1975), Dierker (1982), Bala et al. (1991), and the references cited in these papers.

Introduction: Searching for Paradigms 5

7

Uncertainty

In the last chapter of his book, Debreu (1959) provided a succinct account of how his analysis can be further extended to "the case where uncertain events determine the consumption sets, the production sets, and the resources of the economy." The basic idea9 is that a contract for the transfer of a commodity now specifies, in addition to its physical properties, its location and its date, an event on the occurrence of which the transfer is conditional. This new definition of a commodity allows one to obtain a theory of uncertainty free from any probability concept and formally identical with the theory of certainty. In short, uncertainty is modeled in terms of an expanded set of commodities (with a corresponding expansion of the number of markets), and results on the existence and optimality of an equilibrium parallel to the case of certainty are then obtained with the same analytical apparatus. Of course, taken literally, this expansion of the list of markets raises skepticism about the plausibility of having many participants in all these markets, the usual justification of price-taking behavior. Radner (1968,1970) was the first to provide a comprehensive assessment of the scope and explanatory power of the Arrow-Debreu model of uncertainty.10 First, note that there are some interesting features of the optimization exercise confronting the agents. For a given price system, there is no uncertainty about the value of any production plan, and there is no ambiguity in defining the concept of a profit-maximizing firm. Secondly, given a price system and production plans, there is no uncertainty about an individual consumer /'s net worth (determining f s budget constraint). It follows that beliefs and attitudes towards risk play no role in the assumed behavior of producers (although a consumer fs preference among plans will reflect not only fs tastes but also fs subjective beliefs about the likelihoods of different events and attitudes towards risk). Next, we note that: If the Arrow-Debreu model is given a literal interpretation, then it clearly requires that the economic agents possess capabilities of imagination and calculation that exceed reality by many orders of magnitude. Related to this is the observation that the theory requires in principle a complete system of insurance and futures markets, which appears to be too complex, detailed, and refined to have practical significance. (Radner 1970, p. 457) 9 10

See Arrow (1970, pp. 121-33) for the English translation of his earlier 1953 paper. In what follows in this section, I have drawn directly from Radner's beautiful exposition (1970).

8

Mukul Majumdar

Both of these points - complexity of computation and complexity in the design of institutions for dealing with resource allocation and risk bearing - figure prominently in several chapters of this volume (especially the last three). Limitations on agents' capacities for observation, communication, and computation (as well as other transactions costs) make it intuitively clear why markets are not complete, and hence why there are active but incomplete markets at every date. Once this is made explicit, certain phenomena and institutions not accounted for in the Arrow-Debreu model can be brought within the range of theoretical analysis. In particular, one achieves a better understanding of (i) a theory of demand for money; (ii) stock markets and speculation; (iii) the difficulties of formulating a satisfactory theory of the expected profit maximization by a firm owned partially by agents not involved in management; and (iv) the difficulties of formulating a satisfactory theory of intertemporal consumption and investment of a consumer having an uncertain net worth. Radner's search for paradigms that could enhance the scope of the Arrow-Debreu theory under uncertainty led him to explore a number of equilibrium concepts and to raise new questions about the role of prices in information processing and the efficiency of the market mechanism. The concept of an equilibrium of plans, prices, and price expectations developed by Radner (1972a) became a landmark in the literature on incomplete markets, reviewed by Shafer in Chapter 1. In Radner's view, this definition came closest to the notion of "equilibrium over time" described in Hicks (1939, Chap. X). His study of the existence question led to a new twist: a bound on the size of trades in the asset markets. This was not just an ad hoc element in a particular method of proof, as Hart's example (1975) revealed. Shafer points out that it "makes sense that for markets to remain competitive, no individual agent should participate with a significant share of the market. From this point of view, a priori bounds on the size of trades an individual can make seems reasonable. . . . For many assets, however, there is no such natural bound." This "annoying" property has been one of the themes in a new literature primarily aimed at the existence question. However, the questions of local uniqueness and Pareto efficiency have called for a sophisticated exploration. It is particularly important to remember that an equilibrium need not be locally unique, and may not be Pareto optimal (even viewing models generically). Indeed, there are some subtle issues involving the

Introduction: Searching for Paradigms

9

notion of Pareto optimality, but the robust examples highlight a fundamental break in the link between equilibrium and efficiency as one moves into paradigms of incomplete markets. 6

Heterogeneity of Information

A part of Radner's pioneering paper (1968) dealt with the issue of heterogeneity of information. In the Arrow-Debreu model of uncertainty, each agent has incomplete information about the environment, "but all agents have the same information. This . . . assumption is not tenable if we are to take good account of the effects of uncertainty in an economy." His first task was to construct a formal extension of the model in order to derive a theory of existence and optimality of competitive equilibrium with fixed (but possibly divergent) structures of information. In the context of attaining optimality through markets, Radner noted that "the limits on the information available to the agents . . . and the differences among the information structures of several agents will typically have the effect of drastically reducing the number of markets." First, there is no need for contracts that depend on information that is not available in the economy. Second, the net trade between any group of agents and the rest of the economy can at most depend upon information that is common to both groups (see the example in Radner 1968, Sec. 7). Once one allows for such heterogeneity, yet another concept of "equilibrium" emerges. When traders enter markets with different nonprice information (i.e., information about the economic environment other than prices), there is an opportunity for agents to learn about the environment from prices because market prices "pool," in a possibly complicated manner, the nonprice information signals received by various agents. These inferences are derived, explicitly or implicitly, from individuals' own models of the relationship between the nonprice information received by market participants and the market prices. On the other hand, the true relationship is determined by the individual agents' behavior and hence jointly by their individual models. An equilibrium of this system, in which the individuals' models are identical with the true model, is called a rational expectations equilibrium (see Radner 1967, 1979, 1982). This is the subject of Chapter 2 by Allen and Jordan. The early literature offered little toward understanding any process of information communication and focused primarily on the existence of such an equilibrium - one that reveals all the initial information possessed by all the traders (a positive generic answer was the principal accomplishment of Radner 1979). Subsequent research on efficiency properties is

10

Mukul Majumdar

particularly fascinating, as it enables us to deal more precisely with a number of themes on the "efficiency" of competitive markets in pooling private information (see the fifth section of Chapter 2). "Only for three special parametric classes of utility functions - linear, logarithmic, and exponential - do we find informationally efficient rational expectations equilibria. Moreover, these economies obviously constitute 'knife-edge' examples; they are not generic in any reasonable sense." 7

Learning

In Chapter 3, Blume and Easley review a number of issues in the burgeoning literature on learning, covering single-agent models, dynamic games, and rational expectations equilibria. Economic theorists interested in models of decision making under uncertainty have repeatedly stressed the need to incorporate the process of expectation formation into the "model of an economic man" viewed "as a learning, estimating, searching, information-processing organism" (Simon 1959, Sec. 5; 1972). And it is surely trite to observe that learning is an important element in the improvement of economic efficiency. The theory of stochastic dynamic programming with discounting (Blackwell 1965) has been a basic tool in modeling optimization under uncertainty; the framework allows agents to choose actions in a particular period on the basis of their "experience" or history of the system: the evolution of the states and actions. A notable advance was to extend dynamic programming techniques to the situation where a decision maker has incomplete information about the state (if the "true" state is only partially observable, or if the agent is said to receive a "signal" or "message" about the true state). By explicitly introducing beliefs (as probability distributions over true states or the unobservable components), the decision-making problem was shown (Rhenius 1974) to be logically equivalent to a (Bayesian) dynamic programming problem in which the agent chooses an action on the basis of observed history and beliefs, which are revised or updated according to the rules of the calculus of conditional probabilities. These developments set the stage for a study of the appropriate functional equation of the problem and for deriving some optimal stationary policy. But one faces formidable difficulties in deriving (analytically) any property that can be invoked to offer practical advice. This is true even in the class of deceptively innocuous "bandit" models that have quite concrete structures (see Berry and Fristedt 1985). These models have a particularly attractive feature, since they embody a widespread conflict: acquisition of information has an immediate cost, but may enhance

Introduction: Searching for Paradigms

11

future prospects. However, the numerical computations needed to approximate an optimal policy (calculating the Gittins-Jones index, for example) in even special examples seem to require a combination of programming skills and exceptionally powerful machines that most decision makers do not have at this stage.11 Naturally, analytical complexity is more formidable in the case of a choice of more than one variable and when the decision maker is exposed to more than one source of uncertainty. It is not difficult to understand why we have achieved only modest progress in carrying out the following program of research on the long-run dynamics of an economy suggested by Radner (1970): The evolution of the economy will depend upon the rules or processes of expectation formation and revision used by the agents. In particular, there might be interesting conditions under which such a sequence of momentary equilibria would converge, in some sense, to a (stochastic) steady state. This steady state ... would constitute a second concept of equilibrium. It would take me too far from the central theme of this volume to reflect on the advances in understanding the long-run dynamics of economies. The interested reader can turn to the selective review by Radner (1991), which has an extended list of references. Some of the deterministic, highly aggregated ("match box") Walrasian models - in the tradition of Uzawa (1961a, 1963) - do not deal with optimizing agents explicitly; see Wan (1971) for a survey. But these models do provide insights into questions of convergence of momentary equilibria to a longrun steady state. On the other hand, in Samuelson's (1958) framework of overlapping generations, or in David Gale's "dual" competitive economies supporting optimal allocations, trajectories of momentary equilibria may be chaotic.12 There is certainly room for considerable work on the stochastic steady states of nonlinear dynamical systems that are subject to random shocks. 8

Games and Organizations: Incentives and Efficiency

It is fair to say that, for many writers, the Walrasian mechanism is the canonical model of decentralized decision making in which (Pareto) efficiency is achieved through agents acting in their own self-interest. There is now an impressive literature that challenges the appropriate11

12

At least this is my impression from the admittedly limited look at the relevant literature when I was involved in the preparation of Majumdar and Radner (1993), and from talking to Professors M. El-Gamal and V. Bala. See also Basu, Bose, and Ghosh (1996), which contains an extensive list of references. See the references in Majumdar and Mitra (1994).

12

Mukul Majumdar

ness of utility and profit maximization assumptions from a number of perspectives,13 and the need to examine the issue of "self-interest" in the presence of public goods was raised over forty years ago by Samuelson. If one considers "decentralization" to mean "a special case of division of labor, where the 'labor' in question is that of making decisions,"14 it is natural to study it in the general context of an organization with many decision makers, each having limited information and limited control of all the activities and each acting in self interest as perceived. The outcome to the organization typically depends jointly on these decisions, and perhaps on some stochastic environmental variables. If one recognizes that the rules of organizations are not immutable, then the broad question that one can pose is the following: How can the rules be designed so as to provide agents with proper incentives to make decisions that will promote the objectives (efficiency, long-run survival, equity, etc.) of the organization? Chapter 4 by Radner is an introduction to what is now a vast literature dealing with such issues by using the tools of dynamic games with incomplete information. Again, one is led to the study of equilibria (in the tradition of Cournot-Nash) and their Pareto efficiency properties. In such a framework, two types of problems arise: hidden actions and hidden information (to use the terminology of Arrow 1986), and these are viewed by Radner through "partnership" and "principal-agent" games. A particularly interesting point is that when one considers a relationship that is repeated over time, new opportunities for incentives and efficiency open up because the evolution of the system provides evidence as to the role of hidden elements. An elegant example is provided in Radner (1981), where the law of the iterated logarithm is used in the context of a repeated principal-agent game. A few interesting points that emerge from such a game-theoretic approach have a bearing on some of the mathematical modeling issues listed in Section 2. First, the approach continues to face the same problem of multiplicity of equilibria (here, Cournot-Nash equilibria and many of the subsequent variations), which significantly weakens the incisiveness of prediction (to be sure, repeated or stochastic games do not offer any help in this direction). Secondly, the decentralization of information, when combined with a divergence of interests of the agents, leads to inefficiencies that go beyond what can be attributed to incompleteness of information in itself. Finally, it is not easy to formalize "how things will work themselves out" and how a particular equilibrium is 13 14

A summary of the voluminous literature is in Scherer and Ross (1990). See also Dutta and Radner (1994,1995) and Radner (1996). See Radner (1972b, p. 88).

Introduction: Searching for Paradigms

13

attained (if one thinks of an equilibrium as a rest point of some dynamic process). It should be clear that, in most realistic settings, the organized decision problems are not even approximately repeated. The essay by Dutta and Sundaram (Chapter 5) deals with the question of existence of equilibria in Markovian games. In these games there is a "payoff-relevant" state variable moving through time in response to actions taken in the game (i.e., a stationary stochastic law of motion that, in principle, deoends on the actions of the players). Thus, a player's current actions could affect his future rewards in two ways: (i) through the effect on the evolution of the state variable, and (ii) through the impact on the behavior of other players, who choose their actions on the basis of observed histories of evolutions of states and all actions. One should recall that, in the Arrow-Debreu paper (1954), the existence of a Walrasian equilibrium was formulated as a problem of existence of a Cournot-Nash equilibrium in a "generalized game" - a problem solved by applying a fixed-point theorem.15 But in many Markovian games the existence question turns out to be particularly subtle, and the discussion is of import as these games arise naturally in the analysis of some questions that figure prominently in public-policy debates (for example, the management of renewable and exhaustible resources). I should add here that even in relatively simple formulations, however, the trajectory of states in an equilibrium can behave in a chaotic manner (see Dutta and Sundaram 1993). There is no doubt that the game-theoretic approach leads to a better understanding of "a good many economic relations inexplicable in previously standard analysis," since "contractual relations are frequently a good deal more complicated than the simple models of exchange of commodities and services at fixed prices would suggest" (Arrow 1986). Arrow did note the gap between the complexity of some solutions and their real-life counterparts, and felt that bridging this gap would involve modeling such elements as costs of communication, variety and vagueness of monitoring, and socially mediated rewards - elements that "go beyond the usual boundaries of economic analysis." The last three chapters of the volume can be regarded as attempts to do just that. In Chapter 6, Schotter notes that, although the literature on "mechanisms" inspired by Hurwicz (1972,1986) has become rich in detail, it has tended to remain somewhat abstract and theoretical. His focus is a real15

Subsequent research generated a literature (reviewed in Border 1985) on logical interconnections and equivalences among these equilibrium concepts and fixed-point theorems; see Uzawa (1962a) and Peleg (1967).

14

Mukul Majumdar

world mechanism selection problem that typically confronts the leaders of an organization (e.g., a corporate CEO or government official). Of course, this question of choosing among mechanisms leads to formidable analytical difficulties (one of the earliest contributions is Marschak 1972). Schotter stresses the point that a real-world decision maker may judge mechanisms on the basis of criteria that disqualify many that are "theoretically elegant." A part of the review uses the allocation problem in the baseball industry as a motivating example: here the problem is one of allocating a number of heterogeneous indivisible goods (baseball players) among a set of consumers (baseball teams). Schotter proposes a list of seven criteria that are likely to be appealing to real-world decision makers. "Some of the criteria I have discovered by speaking to corporate leaders, while others I have simply dreamed myself"; it is particularly important to note that equity is one of the criteria (one hopes that the theoretical literature on mechanism design will attempt to incorporate such issues). He goes on to review experimental results, relying on his previous studies and those of Linhart, Nalbantian, Radner, and others. For the baseball industry, three mechanisms are compared, including a laboratory version of the current free-agency system. 9

Bounded Rationality

If one thinks of an individual decision maker "as a learning, estimating, searching, information-processing organism," then it is obvious that management of a large organization will inevitably lead to some form of decentralized decision making, particularly when delays involving decision making prove costly to the organization. "Information processing" includes not only the routine gathering and storage of information but also, and perhaps more importantly, studying the implications of the information and coming to a judgment on the basis of some explicit and/or implicit calculations. Moreover, extending beyond a single individual involves a process of communication or transmission of payoffrelevant information. Within an organization, it is often the case that the process of choosing an action following "a meeting of the managers" involves some or all of these activities. The costs of decision making are considerable: when an adequate performance of these tasks is beyond the capabilities of the individuals responsible, there develops an urge to rely on a supporting "staff" of humans and machines.16 Of course, given 16

Radner (1992) notes that in U.S.firmsa significant fraction of employees - one third to one half - are directly or indirectly engaged in "information processing."

Introduction: Searching for Paradigms

15

decentralization, hidden information, and imperfect monitoring, it is possible for a subgroup of managers to engage in activities that serve themselves better, perhaps in conflict with the overall objective of the organization; the game-theoretic approach is a natural tool for addressing this issue. But this is only one aspect of the study of costs and benefits of decentralization. A second aspect involves looking into the best structure or design of a decentralized organization, and determining endogenously the number and arrangement of information processing agents that can promote efficiency. These and other issues involving returns to scale have figured prominently in the literature surveyed by Van Zandt in Chapter 7. The formal modeling of decentralization has a resemblance to networks of computers, and parts of his essay and the analysis of Mount and Reiter (Chapter 8) explore related ideas.17 Mount and Reiter also face up to the issue of "computational complexity." The motivating applications they present, however, are mostly not to problems of economics, and - as they concede - the value of their particular formalism will eventually depend on whether new economic insights come out of it.

10

Concluding Comments

The models reviewed in the chapters of this volume add to the toolbox of economic theorists and expand the boundaries of economic analysis, or at least leave us with a deeper understanding of the complexities involved in developing theories of households and firms acting under uncertainty. We are in a better position to explain a large number of economic relations that have a direct bearing on the working and efficiency of an economy. On the other hand, some approaches, whether Walrasian or game-theoretic, have yet to resolve the problems posed by bounded rationality and indeterminacy of equilibria. Perhaps there is an inexorable difficulty with these formal models having their emphasis on optimization and consistency among relations, particularly among relations that are not observable or easily refutable. Economics, according to Hicks (1968), surely, is a social science. It is concerned with the operations of human beings, who are not omniscient, and not wholly rational; who (perhaps because they are not wholly rational) have diverse, and not wholly consistent, ends. As such, it cannot be reduced to a pure technics; for we can then say that its concern is with the use that can be made of pure technics by man in society. 17

Radner (1997) pointed out that the idea of representing human bounded rationality with a model of a computer was due to T. A. Marschak and C. B. McGuire (1971).

16

Mukul Majumdar

Can we predict the outcome of events in a "not wholly rational" world through tractable formal models? And, yes, it is not easy for us to offer good advice on how to choose the best action in promoting individual or social welfare. This is understandable: Rushdie (1996) persuades us easily - without formal definitions and theorems - that good advice is rarer than rubies. The Spirit of the Road smiled on Apu and said, "Silly boy, the road never ends . . . it stretches forward from the familiar to the unfamiliar,... I have made you homeless to share the strange pleasures of the journey.... Let us move on." (Bibhuti Banerji in Father Panchali, The Song of the Road) References Arrow, K. J. (1970): Essays in the Theory of Risk Bearing, North-Holland, Amsterdam. (1986): "Agency and the Market," in Handbook of Mathematical Economics, vol. 3 (eds. K. J. Arrow and M. D. Intrilligator), North-Holland, Amsterdam, pp. 1183-95. Arrow, K. J., and G. Debreu (1954): "Existence of Equilibrium for a Competitive Economy," Econometrica, 22, pp. 265-90. Arrow, K. J., and F. Hahn (1971): General Competitive Analysis. Holden-Day, San Francisco. Bala, V., and M. Majumdar (1992): "Chaotic Tatonnement," Economic Theory, 2, pp. 437-45. Bala, V., T. Mitra, and M. Majumdar (1991): "Resource Allocation in NonWalrasian Environments," Journal of Economic Behavior and Organization, 15, pp. 1-28. Basu, A., A. Bose, and J. K. Ghosh (1996): "Sequential Design and Allocation Rules," in Handbook of Sequential Analysis (eds. B. K. Ghosh and P. K. Sen), Marcel Dekker, New York, pp. 475-502. Berry, D. A., and B. Fristedt (1985): Bandit Problems: Sequential Allocation of Experiments, Chapman and Hall, London. Blackwell, D. (1965): "Discounted Dynamic Programming," Annals of Mathematical Statistics, 36, pp. 226-35. Border, K. C. (1985): Fixed Point Theorems with Applications to Economics and Game Theory, Cambridge University Press. Chipman, J. S. (1987): "International Trade," in The New Palgrave (eds. J. Eatwell, M. Milgate, and P. Newman), MacMillan, London, pp. 922-54. Debreu, G. (1959): Theory of Value, Wiley, New York. (1970): "Economies with a Finite Set of Equilibria," Econometrica, 38, pp. 387-92. (1974): "Four Aspects of the Mathematical Theory of Economic Equilibrium," in Proceedings of the International Congress of Mathematicians (Vancouver); reprinted as Chapter 18 in Mathematical Economics: Twenty Papers of Gerard Debreu, Cambridge University Press, 1983.

Introduction: Searching for Paradigms

17

Dierker, E. (1982): "Regular Economies," in Handbook of Mathematical Economics, vol. 2 (eds K. J. Arrow and M. D. Intrilligator), North-Holland, Amsterdam, pp. 795-830. Dorfman, R., P. A. Samuelson, and R. Solow (1958): Linear Programming and Economic Analysis, McGraw-Hill, New York. Dutta, P., and R. Radner (1994): "Profit Maximization and the Market Selection Hypothesis," Working Paper, Stern School of Business, New York University (revised 1997). (1995): "Moral Hazard," in Handbook of Game Theory, vol. 2 (eds. R. J. Aumann and S. Hart), North-Holland, Amsterdam, pp. 869-903. Dutta, P., and R. Sundaram (1993): "How Different Can Strategic Models Be?" Journal of Economic Theory, 60, pp. 42-61. Hahn, F. (1982): "Stability," in Handbook of Mathematical Economics, vol. 2 (eds. K. J. Arrow and M. D. Intriligator), North-Holland, Amsterdam, pp. 744-93. Hart, O. D. (1975): "On the Optimality of Equilibrium When the Market Structure Is Incomplete," Journal of Economic Theory, 11, pp. 418^3. Hicks, J. R. (1939): Value and Capital: An Enquiry into Some of the Fundamental Principles of Economic Theory, 2nd ed., Oxford University Press, London. (1968): "Linear Theory," Economic Journal, 70, pp. 671-709. Hurwicz, L. (1972): "On Informationally Decentralized Systems," in Decision and Organization (eds. C. B. McGuire and R. Radner), North-Holland, Amsterdam. (1986): "On Informational Decentralization and Efficiency in Resource Allocation Mechanisms," in Studies in Mathematical Economics (ed. S. Reiter), MAA Studies in Mathematics, vol. 25, Mathematical Association of America, Washington, DC. Hurwicz, L., R. Radner, and S. Reiter (1975): "A Stochastic Decentralized Resource Allocation Process Parts (i) and (ii)," Econometrica, 43, pp. 187-221 and pp. 363-93. Koopmans,T. C. (1957): Three Essays on the State of Economic Science, McGrawHill, New York. Lange, O. (1942): "The Foundations of Welfare Economies," Econometrica, 10, pp. 215-18. Majumdar, M. (ed.) (1992): Decentralization in Infinite Horizon Economies, Westview, Boulder, CO. Majumdar, M., and T. Mitra (1994): "Some Lessons from the Theory of Chaotic Dynamical Systems: An Overview," presented at the ASSA Annual Meeting (Boston). Majumdar, M., and R. Radner (1993): "When to Switch to a New Technology: Learning about the Learning Curve," AT&T Bell Laboratories, Murray Hill, NX Malinvaud, E. (1953): "Capital Accumulation and Efficient Allocation of Resources," Econometrica, 21, pp. 233-68. (1992): "Foreword," in Decentralization in Infinite Horizon Economies (ed. M. Majumdar), Westview, Boulder, CO. Marschak, T. A. (1972): "Computation in Organizations: The Comparison of

18

Mukul Majumdar

Price Mechanisms and Other Adjustment Processes," in Decision and Organization (eds. C. B. McGuire and R. Radner), North-Holland, Amsterdam, pp. 237-83. Marschak, T. A., and C. B. McGuire (1971): Lecture Notes on Economic Models for Organizational Design, Unpublished manuscript, University of California, Berkeley. Nikaido, H. (1968): Convex Structures and Economic Theory, Academic Press, New York. Peleg, B. (1967): "Equilibrium Points for Open Acyclic Relations," Canadian Journal of Mathematics, 19, pp. 366-9. Radner, R. (1967): "Equilibre des marches a terme et au comptant en cas d'incertitude" Cahiers d'Econometrie (C.N.R.S., Paris) 4, pp. 35-42. (1968): "Competitive Equilibrium under Uncertainty," Econometrica, 36, pp. 31-58. (1970): "Problems in the Theory of Markets under Uncertainty," American Economic Review, 60, pp. 454-60. (1972a): "Existence of Equilibrium of Plans, Prices and Price Expectations in a Sequence of Markets," Econometrica, 40, pp. 289-304. (1972b): "Normative Theories of Organization: An Introduction," in Decision and Organization (eds. C. B. McGuire and R. Radner), North-Holland, Amsterdam, pp. 177-88. (1979): "Rational Expectations Equilibrium: Generic Existence and the Information Revealed by Prices," Econometrica, 47, pp. 655-78. (1981): "Monitoring Cooperative Agreements in a Repeated Principal-Agent Relationship," Econometrica, 49, pp. 1127^8. (1982): "Equilibrium Under Uncertainty," in Handbook of Mathematical Economics, vol. 2 (eds. K. J. Arrow and M. D. Intriligator), North-Holland, Amsterdam, pp. 973-1006. (1991): "Intertemporal General Equilibrium," in Value and Capital: Fifty Years Later (eds. L. W. McKenzie and S. Zamagni), MacMillan, London. (1992): "Hierarchy: The Economics of Managing," Journal of Economic Literature, 30, pp. 1382-1415. (1996): "Economic Survival," Nancy Schwartz Memorial Lecture, Northwestern University, Evanston, IL. (1997): "Bounded Rationality, Indeterminacy and the Managerial Theory of the Firm," in Organizational Decision Making (ed. Z. Shapira), Cambridge University Press, pp. 324-52. Rhenius, D. (1974): "Incomplete Information in Markovian Decision Models," Annals of Statistics, 2, pp. 1327-34. Rushdie, S. (1996): "Good Advice is Rarer than Rubies," in East, West, Vintage, New York. Saari, D. G. (1985): "Iterative Price Mechanism," Econometrica, 53, pp. 1117-31. Samuelson, P. A. (1947): Foundations of Economic Analysis, Harvard University Press, Cambridge, MA. (1958): "An Exact Consumption-Loan Model of Interest With or Without the Social Contrivance of Money," Journal of Political Economy, 66, pp. 467-82.

Introduction: Searching for Paradigms

19

Scarf, H. (1960): "Some Examples of Global Instability of Competitive Equilibrium," International Economic Review, 1, pp. 157-72. Scherer, F. M., and D. Ross (1990): Industrial Market Structure and Economic Performance, 3rd ed., Houghton-Mifflin, Boston. Shafer, W., and H. Sonnenschein (1982): "Market Demand and Excess Demand Functions," in Handbook of Mathemtical Economics, vol. 2 (eds. K. J. Arrow and M. D. Intriligator), North-Holland, Amsterdam, pp. 671-96. Simon, H. A. (1959): "Theories of Decision-Making in Economics and Behavioral Science," American Economic Review, 49, pp. 253-83. (1972): "Theories of Bounded Rationality," in Decision and Organization (eds. C. B. McGuire and R. Radner), North-Holland, Amsterdam. Uzawa, H. (1961a): "On a Two Sector Model of Economic Growth, I," Review of Economic Studies, 29, pp. 40-7. (1961b): "The Stability of Dynamic Processes," Econometrica, 29, pp. 617-31. (1962a): "Walras' Existence Theorem and Brouwer's Fixed Point Theorem," Economic Studies Quarterly, 13, pp. 59-62. (1962b): "On the Stability of Edge worth's Barter Process," International Economic Review, 3, pp. 218-32. (1963): "On a Two Sector Model of Economic Growth, II," Review of Economic Studies, 30, pp. 105-18. Wan, H. Y. (1971): Economic Growth, Harcourt Brace Jovanovich, New York.

CHAPTER 1

Equilibrium with Incomplete Markets in a Sequence Economy Wayne Shafer

I

Introduction

In the now classical extension of the general competitive model to include time and uncertainty, it was observed that if at time 0 (today) there are markets for future delivery of all goods contingent on the date and state of nature, then the resulting equilibrium allocation will be Pareto efficient. Moreover, in such an equilibrium future spot markets would not need to be active; the future would simply consist of agents fulfilling their promises (see Debreu 1959). In the event that there is not a complete system of contingent commodity contracts at time 0, then a new problem arises; there will be reason for future spot markets to be active, so that agents making decisions at the current data need to make expectations about what future prices will be. Thus a new equilibrium concept is required, and that is the topic of this essay. Following the basic paper of Arrow (1964), Roy Radner, in a fundamental series of papers (1967, 1968, 1972), developed the concept of an equilibrium of plans, prices, and price expectations for the sequence of markets that will arise when there is not a complete set of contingent contracts at time 0 - a concept that has led to a substantial amount of research in the last two decades. It is the purpose of this essay in honor of Roy Radner to review this research program that his pioneering work has spawned, and to assess its future. A substantial number of interesting issues arise if all uncertainty is not resolved at the initial date. First, in order to investigate efficiency properties of markets, a new notion of completeness of markets is required, since in this context markets are never complete in the sense of a complete system of contingent contracts available at the initial date. Another development is that a firm's decision problem is now a nontrivial optimization problem in the face of uncertainty, and this has led to an extensive literature on how to model firm behavior. Whether or not returns to assets are denominated in terms of goods or in a unit 20

Incomplete Markets in a Sequence Economy

21

of account becomes important in certain situations, and there is in general no longer a clear dichotomy between the real side of the economy and the financial side, in terms of determining the allocation of resources in equilibrium. We will look at these and other issues in this essay. This paper is organized as follows. In order to have an explicit point of reference, we first describe a canonical model that has become standard in most of the literature on this subject. Then the equilibrium concept is defined, and basic results on existence, determinacy, and efficiency are described. This part of the paper is intended to be an overview, not a comprehensive survey of the area, and in particular we do not discuss the rather substantial technical apparatus that has been developed to deal with this model. For a comprehensive bibliography and a more detailed analysis of the technical problems, the reader is referred to the survey by Magill and Shafer (1991) and the special 1990 issue on incomplete markets of the Journal of Mathematical Economics. The emphasis of this essay is on the conceptual modeling issues that arise, and the final section is devoted to discussing these as well as suggesting some interesting open problems of a conceptual nature. II

A Canonical Model

We begin by describing a canonical model for sequential trading. There will be two time periods, 0 and 1, with S states of nature in time period 1. In each of the two time periods, € commodities are traded, with the trades at time 1 being contingent on the realized state s. Thus the commodity space is 9t€(5+1). We have / > 2 agents, each characterized by a utility function u;. 3?+(++1)^9i and an endowment vector W;E 3t+(++1). Denote by w = (wu . . . , vv7) e dlie(s+l) the list of endowment vectors; this will parameterize the agents. The utility functions are assumed to be smooth, differentiably strictly monotone, and strictly quasiconcave, with the usual boundary conditions. Each agent is supposed to know his own characteristics and each observes the true state when it occurs at time 1. We write a vector y e m€(s+l) in the form y = (y(0),y(l),. . . ,y(S)) with each y(s) e 9i€. A spot price system is a vector p e 3U(++1). We first review the standard competitive equilibrium in this model in which all trades and prices are decided at time 0. We will denote the competitive equilibrium prices by P; P(0) is the vector of prices for goods consumed at period 0, and P(s) (s > 1) is the price vector at time 0 for delivery of goods in period 1 contingent on state s being the realized state. An agent's optimization problem is

22

Wayne Shafer

subject to: ^(o)(x(o) - w,(o)) + £ ^(s)(*(s) - w,(s)) = 0. >l

A competitive equilibrium is a collection ((x;);=i,...,/,P) such that (i) Xt solves agent /'s optimization problem at P; and (ii) 2/(^(5) - wf(j)) = 0 for 5 = 0 , . . . , 5. To introduce the Radner model, we suppose there are / assets which are traded at time 0 and which return dividends at time 1. These are to be viewed as publicly traded assets on competitive markets; this simple model will not include other types of assets that may hedge against risk, such as insurance contracts or labor contracts. A unit of asset j will cost q, units of account payable at time 0 (q = (q^ . . . ,g y )e ?flJ) and return Vj(s) units of account at time 1 in state s. The asset is called nominal if the returns are given exogenously; an asset is called real if the return at state s is the market value of commodity vector, that is, if v;(s) = p(s)cij(s) for a vector afe) e ffl. Of course, mixtures are possible. However, for our simple model we shall consider only the pure real asset case or the pure nominal asset case. One may also consider nonlinear assets such as options, with either real or nominal strike prices. In any case we denote by V the S X / returns matrix that has in row s and column; the dividend of asset j in state s, Vj(s), and let v(s) denote the vector of the returns in state s. In the real asset case, the list of vectors a = (aj(s)) e disa parameterizes the asset structure, and the returns matrix is a function of/?, V(p). For example, suppose each aj(s) is the unit vector with a 1 in coordinate j and 0 elsewhere; that is, asset / promises delivery of one unit of good j in every state. With these commodity forward contracts, v;-(s) = p(s)aj(s) = Pj(s) and V is just the matrix of period-1 prices. In the nominal case the returns matrix V itself parameterizes the asset structure. Denote by Zj the amount purchased of asset /, with z = (zi,... ,Z/)e 9^J. The amount z7 may be positive or negative; the assets are considered to be in zero net supply. We now apply to this model Radner's concept of an equilibrium of plans, prices, and price expectations. At a spot price system p and an asset price system q, define an agent's optimization problem as

subject to: p(oVjt(o) - wf(o)) = -qz, p(s)[x(s) - Wi(s)} = ^ Vj(s)zh s - 1,..., S.

Incomplete Markets in a Sequence Economy

23

The first constraint listed is that the net expenditure on goods plus the cost of the portfolio of assets must sum to zero. The constraints for s > 1 indicate that if s is the realized state in period 1, the net expenditure on goods must equal the dividends of the asset portfolio. The purchase of the assets in period 0 and their dividends in period 1 provides a means both for transferring income between period 0 and period 1 and for transferring income across the potential states in period 1. Note that these constraints preclude the agent from planning bankruptcy in any state; implicit in the constraints is an infinite penalty for bankruptcy. An equilibrium of plans, prices, and price expectations is a list ((xh Z/)/>i, (p, q)) such that: (i) (ii)

(xhZi) solves agent /'s optimization problem at (p,q)', ^(xt{s) - wt(s)) = 0, s = 0 , . . . , S; and

(iii) Igi = 0.

The interpretation of this equilibrium concept is as follows. At time 0, each agent / observes the current spot prices p(0) and the asset prices q, the "prices." Then, based on some beliefs or "price expectations" about spot prices at time 1, say pe{s), s = 1 , . . . ,5, the agent solves the optimization problem, forming the "plans" (xh Zt). If it turns out that the price expectations of every agent are the same and the common expectation p(s) (s = 1 , . . . , S) is such that, together with the observed prices p(0) and q, all markets clear, then we are in an equilibrium. In effect, every agent is correctly predicting the spot prices that would clear the market in each state 5* if that state occurs. We shall discuss this concept critically later on in this essay. This equilibrium concept will also be referred to as a Radner equilibrium. Note that the same concept appears in the literature under other names as well, including "GEI equilibrium," "perfect foresight equilibrium", and "rational expectations equilibrium." Although this is a model with uncertainty, probabilities have not yet been mentioned. Presumably, if agents have Savage-like assessments of probability of states, then these are incorporated into their preferences via their utility functions. The assumption of monotonicity then implies that they give a strictly positive probability to each state, for otherwise they would be indifferent to planning consumption in a state they feel would not occur. However, the model viewed in this way allows agents to have different probability assessments of states. The information requirements of this model are quite strict. As mentioned previously, each agent is fully informed in the sense that he or she will be able to verify the true state once it occurs. In addition, each agent knows exactly the distribution of returns of each asset across the states (i.e., each agent knows each «;(-) in the real asset case and each vy(-) in

24

Wayne Shafer

the nominal case). Thus there is no "inside information." For assets such as corporate stocks, this is a strong requirement. To a great extent, interpretation of both the informational requirements in this model and the equilibrium concept depend on what is meant by a "state of nature," and this will be discussed later. We now define what is meant by completeness of markets in this model. The formal definition is that, for any possible vector of units of account across the S states of nature, an agent can form a portfolio of assets that gives this distribution of returns. That is, for any 5-vector y of net expenditures on goods at period 1 (ys =p{s)(x(s) - w(s))), there is some portfolio z such that y = Vz. In our model, for the nominal case this is equivalent to the returns matrix V having rank = 5, so that the column vectors of V span all of ?fts. In particular, there must be / > S assets. In the case of real assets, there is a problem. The rank of the returns matrix V(p) is a function of p, and is thus endogenous to the model. However, since V(-) is linear in /?, it has a "generic" rank - that is, the rank of V(p) is constant on an open set of full measure in the domain ofp - and this rank is the maximum rank the returns matrix can take on at any p. For the preceding example of commodity forward contracts in which V(p) is the matrix of period-1 spot prices, V(p) is just an arbitrary positive matrix asp ranges over all positive prices, and this just says that most matrices have maximum rank. Thus we say that the real asset structure is complete if this generic rank is S. Again this requires / > S. We note for later reference that if one imposes restrictions on the size of trades an agent can make in the asset market then these markets cannot be complete regardless of the number of assets, since by restricting asset trades z we cannot in general expect to express every vector in $is in the form Vz. We will frequently have to make "generic" statements. If B is a subset of some Euclidean space, then the statement "for almost all b e B" will mean "for all b in some open subset of B whose complement in B has Lebesgue measure 0." A very useful way of developing an understanding for this model, and for the relationship between the Radner model and the classical competitive model, is to evaluate the Kuhn-Tucker conditions (hereafter KTCs) for the agent's optimization problem. We proceed with this now. The KTCs for an agent in the classical competitive model, with Qt the multiplier, are: 5 = 0,..., 5;

Incomplete Markets in a Sequence Economy

25

The KTC for the consumer's problem in the Radner model is as follows, with A,(s) the Lagrangian multipliers for the constraints s = 0, . . . , 5 :

p(s)(xi(s) - TV,(S)) = v(s)zi,

s = 1,..., S.

Suppose we have a competitive equilibrium allocation x with prices P and Lagrangian multipliers 6. In order to obtain this as a Radner equilibrium with the same allocation, one can proceed as follows. Fix positive numbers P(s) (s = 0,.. . ,S) with /3(0) = 1. Define X[(s) = 0fi(s) for all i and s. Definep(s) = l/p(s)P(s), and define q = 2J2:1 P(s)v(s). Then all of the KTCs except possibly the budget constraints will be satisfied. For the budget constraints, choose Zi such that the constraints for 51 = 1 , . . . , S will be satisfied for each agent, and let yt be the S-vector whose s*th coordinate is p(s)(xi(s) — H>/(S)); then these ^-constraints can be written yt = Vzi- Provided that V has rank 5, such a solution will exist (as noted previously), and since Ey, = 0, the Zi can be chosen so that Zz/ = 0. Finally, we need to show that the constraint for period 0 is satisfied. We have f}(s)p(s)(xi(s) - wt(s)) = /3(s)v(s)Zi for s > 1; summing over s gives

Since p(0) = P(0) from the competitive budget constraint, we thus obtain the period-0 constraint for the Radner model. Now suppose we have a Radner equilibrium allocation x with spot prices p, asset prices q, asset trades z, and Lagrangian multipliers X. We will try to construct a competitive equilibrium. First, note that if V has rank S then q = 2iS>1(A/(5')/A/(O))i;(51) has a unique solution given q for the multiplier ratios 2/(5)/Ai(0); in particular, they must be independent of /. (These ratios are sometimes referred to as "personalized state prices"). Thus we can define /3(s) = A/(5)//l/(0) for s > 1 and, as before, let /?(0) = 1. Then define P(s) = P(s)p(s) for s > 0, and set 0t = At(0). With this construction, the KTCs for the competitive consumer will be satisfied except possibly the budget constraint. To show the budget is satisfied also, note that

26

Wayne Shafer

= p(0)(x,(0) =

p(0)(x(0)-wi(0))

Observe that, in both directions, the condition that V have rank S was what allowed us to proceed. In particular, in the second case it guaranteed that the personal state prices A^/A^O) were the same for all agents. As is easily checked, this is precisely the additional condition needed in order to show that if the agents all satisfy the KTCs then we will have equality of all marginal rates of substitution between goods, which is equivalent to Pareto efficiency with our assumptions. In the incomplete markets case, where V has rank less than 5, these personal state prices need no longer be the same. Ill

Existence of Equilibrium

Radner (1972) demonstrated that equilibrium exists in a model more general than the one just described, provided we impose bounds on the size of trades in the asset markets. (The size of the bound is irrelevant; it is necessary only that there be one.) Theorem 1. In the model described in Section II, if each agent's optimization problem is modified with an additional constraint of the form z>h, then an equilibrium of plans, prices, and price expectations exists. It makes sense that, for markets to remain competitive, no individual agent should participate with a significant share of the market. From this point of view, a priori bounds on the size of trades an individual can make seems reasonable. However, the same reasoning applies to the commodity markets, and no such bounds are necessary in this case. In fact, one could easily impose a bound on the size of trade in commodities without affecting the equilibrium; simply make the bound larger than the aggregate supply. For many assets, however, there is no such natural bound. Indeed, even for contracts for forward delivery of commodities, as long as they can be satisfied by payment in units of account of the market value of the contract, there is no requirement connecting the amount of forward trades with the actual supply of the good. Perhaps this explains a rather annoying property of the model; as Hart (1975) discovered, without exogenous bounds on the size of trades in the asset markets, equilibria may fail to exist. Put another way, if one imposes a

Incomplete Markets in a Sequence Economy

27

bound on asset trades, then the bound may be binding in equilibrium no matter how large the bound. One might argue that if the data of the economy are such that agents want to make very large trades, perhaps the data are inconsistent with the hypothesis of competitive behavior. However, Hart's example can be embedded in a model with a nonatomic measure space of agents, where each agent's trades will indeed be negligible, and the same nonexistence problem will arise. This problem of nonexistence when bounds are not imposed on the size of trades has lead, in the last fifteen years, to a substantial number of papers designed to resolve this problem. The main characteristic of Hart's example is that the returns matrix changes rank with /?, and one approach has been to restrict attention to asset structures that do not exhibit this behavior. Cass (1984) and Werner (1985) showed that, if one restricts attention to pure nominal asset structures, then equilibria exist without imposing bounds. Similarly, Geanakoplos and Polemarchakis (1986) observed, in the real asset case, that if all assets are denominated in terms of the market value of a single good then the returns matrix V(p) will have constant rank and - just as in the nominal case - equilibria always exist. In general, then, we have the following theorem. Theorem 2. In the model described in Section II, if the asset structure is such that the matrix of returns has constant rank, then an equilibrium of plans, prices, and price expectations exists. The mere presence of two assets that are commodity forward contracts will in general preclude the hypothesis of Theorem 2. In this case the columns of V(p) corresponding to the two assets will just be the vector of prices for these goods across the states, and these will inevitably be collinear for some prices. In effect they become perfect substitutes, and the returns matrix will drop rank (unless there are at least 5 - 1 other independent assets). This results in a failure of upper hemicontinuity of the demand correspondence defined by the constraints of the agents' optimization problems, and this is the technical reason for nonexistence. In this case the best that one could hope for is that, in some reasonable sense, equilibria exist generically. To illustrate how equilibria may fail to exist, we provide a simple example. Consider the Radner model with one state in period 1, one asset, two consumers, and two goods, with the following data. Endowments are wfe) = (1,1) for s = 0,1. Utility functions are ux(x) = Vi(x(0)) + VI(JC(1)) with v1(x) = (l/3)tnxl + (2/3)€nx2 for agent 1 and ui(x) = v2(x(0)) + (1/2)V2(JC(1)) with v2(x) = (2/3)€n*i + (l/3)€njt 2 for agent 2. The competitive equilibrium prices can be easily computed in

28

Wayne Shafer

this log-linear economy; they are ^(0) = 11/36, P2(0) = 10/36, P^l) = 7/36, and Pi(X) - 8/36. Now consider a Radner version of the model, with one real asset given by ax = (8, -7). The return on this asset in state 1 is V = p1(l)S -p 2 (l)7, so investing in this asset is essentially a bet that the relative price of good 1 in terms of good 2 is greater than 7/8. Since there is one state and one asset, this is the complete markets case. Note that the 1 X 1 returns matrix drops rank precisely in the case the relative price is 7/8 in state 1. We now show that this Radner model does not have an equilibrium. First, we try for an equilibrium with the return V =t 0. If such an equilibrium existed, it would have to coincide with the competitive equilibrium since V has rank 1, but in the unique competitive equilibrium the period1 price ratio is 7/8 and so V = 0, a contradiction. Second, consider the possibility of an equilibrium with V=0. Then there would be no transfers of income between periods 0 and 1, so the period-1 equilibrium would have to coincide with the static competitive equilibrium with the utility functions vx and v2. But it is easy to see from the symmetry of the functions and the equal endowments that the relative price ratio in this case would be 1, and thus V =£ 0, again a contradiction. Thus no equilibrium exists. Note that this example is not robust: Alter the asset a small amount so that V ¥= 0 at the price ratio 7/8 and then the complete markets case will work; or alter endowments or utility parameters a little so that the period-1 price ratio is no longer 7/8, and the complete markets case again works. This example gives a clue on how to proceed for the existence problem with real assets when markets are complete as defined in Section II. In this case, remember, V(p) has constant rank S on an open set of full measure in the space of prices. As we have demonstrated using the KTCs, an equilibrium of plans, prices, and price expectations at which V(p) has rank S is equivalent to a competitive equilibrium with a complete set of contingent commodity contracts. That is, the allocations will be the same, and there is an easy correspondence between competitive equilibrium prices and prices in a Radner equilibrium. Thus a natural strategy in this complete markets case is first to obtain a competitive equilibrium, which always exists in our model, and then to construct the corresponding Radner prices. If, at these prices, V(p) has rank S, then we have an equilibrium of plans, prices, and price expectations. Kreps (1982) made the critical observation that if the rank of the returns matrix is less than 5 at such prices, a small perturbation of the returns structure a will restore V(p) to full rank (as in our preceding example) and thus we have an equilibrium. That is, generically in a, a Radner equilibrium exists. Similarly, Magill and Shafer (1990) observed that a small perturbation

Incomplete Markets in a Sequence Economy

29

of endowments would cause the competitive equilibrium prices to move in the region where V(p) has full rank, and thus equilibrium exists generically in endowments. Theorem 3. In the model described in Section II: / / J > S then, for each w e ?ftiI(++1), an equilibrium of plans, prices, and price expectations exists for almost all a in ?Hm; (2) for each asset structure a for which V(-) has generic rank S, an equilibrium of plans, prices, and price expectations exists for almost all endowment lists w in (1)

In the case where both V(p) can change rank withp and markets are not complete (in particular, if J < S), the trick of first obtaining a competitive equilibrium and then converting it to a Radner equilibrium is no longer available. Nevertheless, by defining a "pseudo" equilibrium concept that replaces the competitive equilibrium in the argument for the complete market case, Duffie and Shafer (1985) were able to show that a Radner equilibrium exists generically in both a and w. Theorem 4. In the model described in Section II with all real assets, an equilibrium of plans, prices, and price expectations exists for almost all

This result is not as satisfactory as the complete market case, because the sense of genericity is weaker. It is still an open question, for example, whether or not equilibrium exists generically in w for any fixed asset structure. Bottazi (1995) has recently shown that this is true, for "most" asset structures, and provides criteria for checking a particular asset structure for this property. IV

Determinacy

We now look at the issue of local uniqueness. In what follows, in "counting" equilibria we are counting the number of equilibrium allocations, since there are certain redundancies in equilibrium prices. We now look at the issue of local uniqueness. In the complete market case this is fairly straightforward, requiring only an adaptation of Debreu's (1970) seminal argument, since competitive equilibria and Radner equilibria coincide when the returns matrix has full rank. In the case of incomplete markets and all real assets, an argument similar to Debreu's applied to the "pseudo" equilibrium also works. The only difficult part of the following theorem is in the real asset case, showing that generically less than maximal rank equilibria do not exist.

30

Wayne Shafer

Theorem 5. In the model described in Section II, if the asset structure is such that markets are complete, or if all assets are real, then for almost all w in the space of endowment lists there exist a finite number of equilibria of plans, prices, and price expectations, and each equilibrium is locally a smooth function of endowment lists w and asset structures a or V. The situation with nominal assets and incomplete markets is completely different, and leads to one of the most fascinating and controversial properties of the Radner model. The following result was proven by Geanakoplos and Mas-Colell (1989) and a similar result by Balasko and Cass (1989). Theorem 6. In the model of Section II with nominal assets, let the returns matrix V be in general position, J <S, and I > J. Then, for almost all w in the space of endowment lists, the set of allocations of an equilibrium of plans, prices, and price expectations contains a set homeomorphic to In order to understand why this result holds, consider first the real asset case. Since each Vj(s) =p(s)aj(s), the constraints in the agent optimization problem are homogeneous in prices. Thus one may separately normalize the period-0 prices (p(0),q) and each period-l price p(s) without affecting the constraints - for example, by picking good 1 to be the numeraire in every state. To most easily see this and what happens in the nominal case, let us look specifically at the agents' constraints for the one-good, two-state case, withp(O) = 1 as a normalization:

~

-*•»

•

•

•

'

Note that, in the case of all real assets where Vj(s)= p(s)aj(s), the period-l constraints take the form

{s) -

s = 1,..., 5,

and so the period-l prices disappear completely (in the multi-good case, this corresponds to only relative prices in each state matter, not absolute price levels). As seen in the KTC analysis, if markets are complete, then the resulting equilibrium allocations will coincide with the competitive allocations, and in particular will be independent of the particular afo). Suppose markets are incomplete, with (say) one asset and two states.

Incomplete Markets in a Sequence Economy

31

Then, from the budget constraints alone, the period-1 consumption vector is constrained to satisfy

for some value of zu and this obviously is a severe constraint on what equilibrium allocations can be. Moreover, changing the relative values of ax(\) and ax{2) necessarily changes the allocation if there is any trade at all; this, in contrast to the complete markets case, implies that the exact structure of the assets now influences the equilibrium allocation. Now consider the nominal case, where the v;-(s) are denominated in the unit of account. Again, the equilibrium allocations in the complete markets case will correspond to the competitive allocations. In the incomplete markets case, however, matters are significantly different. Consider again the one-asset, two-state case, and simply define ax{s) by fixing p{s) arbitrarily positive and setting ax{s) — vx{s)lp(s). Then, for these fixed period-1 prices, the corresponding real model defined will have an equilibrium. Now, if we change the relative values of the p{s), we obtain a new model with different relative ax{s) and thus a different equilibrium allocation. In the two-state case there is a one-dimensional set of relative period-1 prices, so there will be a one-dimensional set of equilibrium allocations. In other words, in this case only the common expectations of period-1 prices in a Radner equilibrium really matter and are not constrained by the equilibrium concept itself. For the general /-good case, the foregoing discussion can be interpreted as saying that one may arbitrarily fix the price level in each state, and for different relative price levels one obtains different equilibria. For example, choosing good 2 as a numeraire instead of good 1 will in general alter the resulting equilibrium. Or, with fiat money and a simple quantity equation to determine the price levels, altering the money supply across states will alter the equilibria. In any event, the term "unit of account" now becomes critical, and equilibria are not determinant unless the agents are given extra information so they can correctly predict price levels in each state. The notion of correct price expectations in the Radner equilibrium is not sufficient by itself to peg price levels in each state, but is only sufficient for pegging relative prices within each state. This leads to the conclusion that the monetary or banking system, which determines the unit of account and its purchasing power, needs to be explicitly incorporated into the model as a necessary condition for obtaining determinacy. (Whether it will be sufficient may depend on the

32

Wayne Shafer

structure of the monetary system; for example, in the simple-minded quantity theory mentioned previously, it would require that the monetary authority announce at time 0 what the supply of fiat money would be in period 1, contingent on the state, or at least that agents could infer it from the observed behavior of the authorities. It is not at all clear that this would be a reasonable hypothesis.) Magill and Quinzii (1992) have examined one such interesting model; much more work needs to be done. In a general sense, the model with an incomplete nominal asset structure provides a framework in which financial institutions can have real effects even in the absence of asymmetric information. Y

Efficiency and Comparative Statics

As mentioned in Section III, at a Radner equilibrium at which the returns matrix has rank 5, the resulting equilibrium allocation will also be a competitive equilibrium allocation, and thus fully Pareto efficient. Thus we have our next theorem. Theorem 7. For the model described in Section II: (1) in the nominal asset case, if\ has rank S then every Radner equilibrium allocation is Pareto efficient; (2) in the real asset case, if the asset structure is such that the generic rank of V(9) is S then, for almost all w in the space of endowment lists, Radner equilibrium allocations are Pareto efficient. In case of incomplete markets, one certainly does not expect full allocative efficiency, for otherwise our definition of completeness would be suspect. The following result indicates this. Theorem 8. For the model described in Section II, ifJ<S then for (w, a) in an open set of full measure in ?Hm X 3t+(s+i;, Radner equilibrium allocations are not Pareto efficient. One might hope that, in some suitable sense, Radner equilibria are constrained efficient. There is still no generally accepted notion of what the correct definition of "constrained efficiency" might be in this case; some argue that the concept cannot be properly defined unless the reasons for incompleteness of markets are endogenously embedded into the model. Nevertheless, we can discuss certain efficiency properties of the equilibria. First, there are robust examples of the model with multiple equilibria in which two of the Radner equilibria have the property that one Pareto dominates the second. One such example is as follows. Consider a simple two-good, two-consumer, one-period economy (an

Incomplete Markets in a Sequence Economy

33

Edge worth box). Suppose preferences are such that there are three equilibria, which we call A, B, and C for simplicity, with a strictly convex contract curve and relative good-1 equilibrium prices pa, pb, and pc. The equilibria are labeled so that ux(A) < U\(B) < ux{C) and u2(A) > u2(B) > u2(C) for the strictly concave utility functions ux and u2. In these circumstances, the line segment joining A and C in the Edge worth box must have an open region where it lies below the indifference curves through B from the perspective of both agents; suppose 0.5A + 0.5C is one such point. Now consider a Radner model with one state in period 1 and utility functions Ui(x(0)) + Ui(x(l)) for each agent. Suppose there are no assets, so that markets are incomplete. Then, for example, BB (B in period 0, B in period 1) is an equilibrium, as is AC. But u^A) 4- ut(C) < 2ui(0.5A + 0.5C) pa and that agents believe s = 1 signals high prices and s = 2 signals low prices. Then the price expectations of pc in state 1 and/?! in state 2 can be realized as a Radner equilibrium with equilibrium allocation C in state 1 and A in state 2. Note that, since there are now two states and still no assets, this is an incomplete market equilibrium. Moreover, if the agents believe the states have equal probability, then 0.5ui(A) + 0.5^(C) < Ui(0.5A + 0.5C) < ut(B), so that the equilibrium of C in state 1 and A in state 2 is Pareto dominated by the equilibrium with B in both states. With these beliefs about states mattering, suppose an entrepreneur sees that introducing an asset to accomodate transfers of income across the two states would be welcome, and does so. Then we would have a complete market model, and the equilibrium allocations would have to be AA, BB, or CC across the two states. But in any of these equilibria there would be no need to trade in the asset, so the demand for the asset would be zero. If the entrepreneur correctly anticipated this demand, she would have no incentive to introduce the asset in the first place. Thus exogenous states that may not be relevant in an intrinsic sense may still have an effect on the Radner equilibria of the model, and can change the market structure from complete to incomplete. Comparing the number of assets with the number of states seems a silly exercise, so one possibility is to alter the definition of completeness. One would like the dichotomy between complete and incomplete markets to coincide with the dichotomy of Pareto efficiency and inefficiency; the definition employed here is in one sense too strong. With our definition of completeness of the asset market, Radner equilibrium will be efficient regardless of the structure of preferences and endowments. It seems unlikely, however, that the observed asset structure arose independently from the basic data defining the economy. A more sensible definition of completeness perhaps should involve a joint condition on assets and on preferences and endowments. For example, in the standard one-good capital asset pricing model in which agent utilities depend only on the mean and variance of a portfolio, only two independent assets are required in order to achieve efficiency in a Radner equilibrium, regardless of the number of states. An alternate approach to defining completeness - one not directly based on the number of assets versus the number of states - is the idea that some agents may be restricted from participating (or participating

36

Wayne Shafer

fully) in the markets for some assets. Such restrictions, for example, could arise if some agents could not verify states upon which the asset payment is conditional. See Balasko, Cass, and Siconolfi (1990) for such an analysis applied to the model with nominal assets. The foregoing approach of restricted participation may be viewed as an initial attempt at studying the issue of why markets may be incomplete, whatever the definition. One needs to emphasize that the model considered here is appropriate for describing publicly traded securities, and these markets alone cannot be expected to handle all types of risk. Insurance markets handle different types of risk and there are adequate theoretical explanations (e.g., moral hazard) of why such contracts are not optimal for such risks. Our model excludes such risks; there is no moral hazard or asymmetric information in the model, for example, and one may reasonably ask why such markets may be incomplete without introducing such issues. The basic argument goes like this: If there are risks for which no existing portfolio of assets can hedge, then entrepreneurs will see this and introduce such new assets. (But remember the sunspot example, in which a new asset will not be traded in equilibrium.) In fact, new securities are issued regularly on publicly traded markets. Most of these are portfolios of existing assets, usually requiring trading strategies too complex to be carried out by individual investors. In fact, when looking at a multi-period version of our model with a maximum of (say) K branches at any node in the event tree, our version of completeness requires only K independent securities (see Kreps 1982). But the required portfolios or trading strategies can be quite complex, and one natural source of restricted participation is that most agents may not be able to implement the more complicated trading strategies, or even discover what they are. Thus computational complexity may be a source of incompleteness. Returning to newly introduced securities, occasionally they are really new, such as a futures contract on a commodity for which no such contracts previously existed. This is both evidence that entrepreneurs are looking for opportunities to introduce new assets and evidence that markets are not yet complete - that is, there are risks which can be handled by a publicly traded security and for which no such security exists. As yet we have no suitably general model of asset introduction that permits the asset structure to become endogenous (the closest is the work by Allen and Gale 1988), and until we do we cannot answer the question of whether such a model would lead to complete markets. One major source of publicly traded assets are shares and bonds issued by corporate firms, which are absent from our canonical model. There are two main difficulties with introducing firms into the model, one con-

Incomplete Markets in a Sequence Economy

37

ceptual and one technical. The conceptual issue is how to model the firm's objective without leaving the competitive markets framework. This is straightforward for a single-owner firm but a nontrivial problem for a corporate firm. To see this, we construct a simple example of the Radner model with a corporate firm. Take the case of one good and no assets except for the shares of a single firm. The firm uses the good in period 0 as an input (y(0)), and obtains a state-dependent output in period l,y(s) for s > 1 according to a production relation f(y) = 0. Agent / is initially endowed with share yt of the firm at the beginning of period 0, and her demand for share ownership in period 0 is denoted by ax. In this case the agent optimization problem is

= maxui(x) Xfii

subject to: x(o) - W;(o) + [q + y(o))a = qy, x(s) — Wi{s\ = y(s)a,

s = 1,..., S.

Here Ut denotes an indirect utility function depending on the firm decision y. By Shephard's lemma, dUi(y)/dy(0) = -cd;(0) and dUi(y)/dy(s) = aXt{s) for s > 1, where the At(s) are multipliers for the agent optimization problem. Define fii(s) = A/(.s)/A/(0), the "personal state prices" as in the KTC analysis of Section II. Note that, with this notation, one of the KTCs y Es^i/3i(s)y(s). for the optimization problem is q = —y(0) + The first-order conditions for the following problem of the agent choosing what she would like the firm to do,

subject to: f(y) = 0, are fii(s) = -df(y)/dy(s) for s > 1. Now consider the general problem

subject to: f(y) = 0. The KTC for this problem, eliminating the multiplier, is /3(s) = -df(y)/dy(s). This means that agent / would like the firm to maximize a weighted sum of dividends, the "market value," using her personal state prices as the weights. Thus, if the personal state prices /?,- are the same for all / (the complete markets case), then all agents agree that the firm should maximize the market value according to the common /?. However, if the /?/ are different then the agents want different objective functions for the firm (see Radner 1974).

38

Wayne Shafer

A number of proposals have been made to overcome this problem. One type of proposal is for the firm to use, as state prices in its objective function, a weighted average of shareholders' personal state prices across the S states (these will be generically different if markets are incomplete). These weights could be the posttrade fractions of shares held (Dreze 1974) or the initial fractions of shares held (Grossman and Hart 1979). Both papers give arguments to the reasonableness of their proposal, and one may argue about which is better, but in the end the choice is somewhat arbitrary. Regardless of how one chooses state prices for the firm to use, the choice will affect the firm's behavior and thus the equilibrium allocations that come about. Since there seems to be no natural way to define a competitive objective function for a corporate firm when markets are incomplete, it may be most appropriate to model the firm as a strategic player. Even if the data of the model are such that it is a reasonable approximation to suppose the firm behaves as a perfect competitor in the commodity markets, it is another matter to suppose that it is a perfect competitor in the asset markets. The shares issued by the firm need have no perfect substitutes if the asset markets are incomplete, and the firm should recognize that its market value will depend not only on its own production decision but also on the production decisions of all other corporate firms. Modulo difficulties associated with multiple equilibria, one can readily describe the market value of each firm as a function of the decisions of allfirms,and in this context a Nash equilibrium seems appropriate. There are, however, serious technical difficulties associated with such a setup, and it is not clear if such a research program can be successfully carried out. Even if one chooses one of the competitive objective functions, there is a serious technical problem associated with production sets. Existing theorems providing existence of Radner equilibria impose strong smoothness assumptions on production sets, and such smoothness assumptions are absolutely essential. Consider the simple case of a constant-returns technology with a capacity constraint. There will typically be an open set of prices p at which the firm would have negative market value if it chose to produce, and so chooses not to produce at all, and another open set of prices at which the firm will choose to produce and its shares will have positive market value. In effect, the number of assets available will be a function of p, and in this example the corresponding returns matrix will no longer have a generic rank. This precludes any generic existence theorem. Finally, we look at the notion of an equilibrium of plans, prices, and price expectations in more detail. One may seriously question whether

Incomplete Markets in a Sequence Economy

39

this is a reasonable solution concept; the requirements are really quite strong. When states of nature are interpreted as exogenous events affecting the data of the economy, the requirement of common and correct price expectations seems inordinately strong. In a literal sense, it requires the agent to know all the data of the economy - not just his own characteristics - and to compute the general equilibrium. However, one may also view the solution concept as an idealization of the idea that agents active on the markets for assets learn the relationship between states and prices. For example, if the state of nature represents adverse conditions for growing a commodity, thus predicting a small harvest, agents understand that the price will be higher than normal. In other words, there is certainly knowledge gained from experience about the relationship between states of nature and prices, and the Radner equilibrium concept recognizes this. It is, however, a great leap of faith to suppose that agents know the relationship exactly. On the other hand, the notion of temporary equilibrium does not suppose the agents know anything about how future prices are related to states of nature, and this seems too extreme in the opposite direction. A more reasonable solution concept would seem to be one in which agents are presumed to have partial knowledge about future prices, and I think this is a useful direction for future research. It is important to remember, and a positive aspect of the equilibrium concept, that agents are not required to have the same probability beliefs about states of nature. That is, one agent may view a particular state as having a small probability of occurring, while another believes the same state has a high probability of occurrence. One way to seemingly avoid the assumption that agents knowp(«) is to view a state of nature as including endogenous outcomes of the model. One could, for example, define a state of nature s in the form s = (e,p, Wi,..., W/), where e represents an exogenous observable event,/? a price vector for the € commodities, and wt an endowment /-vector for agent /. In this case the period-1 price function is just the appropriate projection map on this new set of states and hence trivial to compute. Likewise, the agent's period-1 endowment will be a projection map and hence trivial. It would appear that with this interpretation the strong informational requirement of the Radner equilibria disappears. However, the problem has simply shifted. In this setup, the agents must form preferences over consumption plans contingent on states, and we may suppose that agents achieve this by forming subjective probability measures over states with a common support and plans that are consistent. For example, consider our sunspot example with one relevant state e. Enlarge the state space to the new definition. Then the equilibrium B, for example, would be the

40

Wayne Shafer

outcome if the subjective probability distributions of the agents give probability 1 to the state s = (e,pb, wu ••.,W/). However, if their subjective probabilities assign 0.5 to the state (e,p a ,w u .. . ,w7) and 0.5 to the state (e,pc,wl9.. . ,w7) then the sunspot equilibrium will be the outcome. Thus, defining the state space in this new way - so as to avoid the informational requirements of the Radner equilibrium - introduces all possible sunspot equilibria as well as the original equilibria. And since agents must have similar probability assessments in the sense of agreeing which states have positive probability, it does not remove the "common" requirement from the expectations assumption of a Radner equilibrium. Whatever one's view of the Radner equilibrium concept, it is difficult to conceive of how one could change it and still have a framework for studying problems involving the efficient allocation of resources under uncertainty. It seems to be the correct concept in this context, in particular because it yields outcomes equivalent to the standard competitive equilibrium with complete contingent contracts when markets are complete in a natural way. Radner's equilibrium concept is also central to the notion of no arbitrage pricing of assets; without the hypothesis that agents have common price expectations, agents will not have common beliefs about the returns of real assets and thus will have different views about the presence of arbitrage opportunities. References Allen, E, and D. Gale (1988), "Optimal security design," Review of Financial Studies 1: 229-63. Arrow, K. (1964), "The role of securities in the optimal allocation of risk bearing," Review of Economic Studies 31: 91-6. Balasko, Y., and D. Cass (1989), "The structure of financial equilibrium: I. Exogenous yields and unrestricted participation," Econometrica 57:135-62. Balasko, Y., D. Cass, and P. Siconolfi (1990), "The structure of financial equilibrium with exogenous yields: II. Endogenous yields and restricted participation," Journal of Mathematical Economics 19:195-216. Bottazi, J. (1995), "Existence of equilibrium with incomplete markets: The case of smooth returns," Journal of Mathematical Economics 24: 59-72. Cass, D. (1984), "Competitive equilibria in incomplete financial markets," Working Paper no. 84-09, CARESS, University of Pennsylvania, Philadelphia. Debreu, G. (1959), Theory of Value, New York: Wiley. (1970), "Economies with a finite set of equilibria," Econometrica 38: 38792. Dreze, J. (1974), "Investment under private ownership: Optimality, equilibrium,

Incomplete Markets in a Sequence Economy

41

and stability," in J. Dreze (ed.), Allocation under Uncertainty: Equilibrium and Optimality. New York: Wiley, pp. 129-65. Duffie, D., and W. Shafer (1985), "Equilibrium in incomplete markets I: A basic model of generic existence," Journal of Mathematical Economics 14: 185-200. Geanakoplos, J., and A. Mas-Colell (1989), "Real indeterminacy with financial assets," Journal of Economic Theory 47: 22-38. Geanakoplos, J., and H. Polemarchakis (1986), "Existence, regularity, and constrained suboptimality of competitive allocations when markets are incomplete," in W. P. Heller, R. M. Starr, and D. A. Starett (eds.), Uncertainty, Information and Communication: Essays in Honor of Kenneth Arrow, vol. 3. Cambridge University Press. Grossman, S., and O. Hart (1979), "A theory of competitive equilibrium in stock market economies," Econometrica 47: 293-330. Hart, O. (1975), "On the optimality of equilibrium when the market structure is incomplete," Journal of Economic Theory 11: 418-43. Kreps, D. (1982), "Multiperiod securities and the efficient allocation of risk: A comment on the Black-Scholes option pricing model," in J. McCall (ed.), The Economics of Uncertainty and Information. University of Chicago Press. Magill, M., and W. Shafer (1990), "Characterization of generically complete real asset structures," Journal of Mathematical Economics 19:167-94. (1991), "Incomplete markets," in W. Hildenbrand and H. Sonnenschein (eds.), Handbook of Mathematical Economics, vol. IV. New York: Elsevier, pp. 1523-614. Magill M., and M. Quinzii (1992), "Real effects of money in general equilibrium," Journal of Mathematical Economics 21: 301^2. Radner, R. (1967), "Equilibre des marches a terme et au comptant en cas d'incertitude," Cahiers d'Econometrie (C.N.R.S., Paris) 4: 35-42. (1968), "Competitive equilibrium under uncertainty," Econometrica 36:31-58. (1972), "Existence of equilibrium of plans, prices, and price expectations in a sequence of markets," Econometrica 40: 289-303. (1974), "A note on unanimity of stockholder's preferences among alternative production plans: A reformulation of the Ekern-Wilson model," Bell Journal of Economics and Management Science 5:181-4. Werner, J. (1985), "Equilibrium in economies with incomplete financial markets," Journal of Economic Theory 36: 110-19.

CHAPTER 2

The Existence of Rational Expectations Equilibrium: A Retrospective Beth Allen and James S. Jordan

1

Introduction

The term "rational expectations" was introduced in the famous article of Muth (1961), but the equilibrium existence problem posed by such a concept had been recognized and addressed formally seven years earlier. Grunberg and Modigliani (1954) recognized that predictions of economic events, unlike weather forecasts, can affect predicted events. Expectations of a substantial price increase, for example, can trigger investment and production decisions that increase supply and cause the actual price increase to be smaller than expected. Weather forecasting may be difficult in practice, but even the logical possibility of economic forecasting is problematic. Although they did not model the process of price determination explicitly, Grunberg and Modigliani showed that Brouwer's fixed-point theorem implies the existence of a correct forecast, provided that the requisite continuity and boundary conditions are satisfied. Muth (1961) acknowledged the Grunberg and Modigliani article, but focused more on using the assumption of rational expectations to specify expectations in econometric models in which the equilibrium could be obtained by construction. The dramatic advances in the mathematical theory of general equilibrium during the fifties and early sixties stimulated efforts to include expectations and uncertainty in general equilibrium models during the late sixties and early seventies. Two articles by Radner (1968 and 1972) stand out from the many notable contributions of this period for their explicit focus on the information structure underlying the formation of Preparation of this manuscript was supported by NSF Grants SBR-9309854 and IRI9312783. Previous financial support from the National Science Foundation for much of the research described here is also gratefully acknowledged. The views expressed herein are those of the authors and not necessarily those of the Federal Reserve Bank of Minneapolis or the Federal Reserve System. 42

Rational Expectations Equilibrium: A Retrospective

43

expectations. Radner's models imposed the strong informational consistency requirement that an agent's actions could not differ across states of the world that could not be distinguished by the agent's information. This condition restricts trade among agents to events that are mutually observable. As Radner (1968, p. 50) emphasized, "the net trade between any group of agents and the group of all other agents in the economy can at most depend upon information that is common to both groups of agents" [italics in original]. As a result, Radner's existence theorems relied on the assumption that all agents were endowed with a common information structure. The possibility that information revealed by prices might remove incongruities in initial information was not pursued. In an earlier unpublished technical report, Radner (1967) had explicitly added the information revealed by prices to agents' information structures. However, the discussion of the existence of equilibrium was confined to the prophetic remark: "It is not clear that an equilibrium of the type defined above exists, even under the classical assumptions. The continuity of the demand functions can, in principle, be destroyed by the fact that information depends on the structure of spot prices." The first positive and negative results on price-conditional rational expectations equilibrium were both achieved by Jerry Green (1973 and 1977, respectively). Green (1973) described a market for the exchange of state-contingent wealth claims in which some traders are privately informed of the "true" state probabilities, while other traders have no private information. Green showed that the traders' excess demand functions, when differentiated with respect to the state probabilities, have a dominant diagonal property that implies that the function from state probabilities to market-clearing prices is one-to-one. Hence there exists in this model a rational expectations equilibrium in which market prices fully reveal the state probabilities to the uninformed traders. Grossman (1981) and Handel (1980) later showed that the key property of contingent claim demand is sufficiently general that the differentiability of demand functions is unnecessary for the result. However, Green (1975, later published as Green 1977) showed that the absence of "noise" in the model is essential, not only for the full revelation property, but even for the existence of equilibrium. More specifically, Green constructed an example in which traders' endowments are subject to state-independent random noise, and showed that the example admits no rational expectations equilibrium. The contingent claims market setting of Green's counterexample made the example necessarily complex. Subsequent authors, notably Kreps (1977), discovered much simpler examples of the nonexistence of equilibrium in spot market models. One such example will be described

44

Beth Allen and James S. Jordan

in the next section. Thus, while the causal influence of economic expectations that troubled Grunberg and Modigliani can be resolved, as they suggested, by conventional fixed-point methods, the informational discontinuity discovered by Radner can be fatal to the existence of rational expectations equilibrium even under classical assumptions on trader characteristics. What follows is an exposition and review of selected aspects of the rational expectations equilibrium existence problem and its literature. We make no pretense of completeness, but refer the reader to Radner (1982) for additional references and topics. Section 2 defines the conventional concept of rational expectations equilibrium (REE), and Section 3 contains a simple example in which the equilibrium fails to exist. Section 3 also reviews the literature on the generic existence of REE. Section 4 relates the existence problem to an earlier model of informationally decentralized allocation mechanisms, and demonstrates the general existence of equilibrium when expectations are conditioned on trades as well as prices. Sections 5 and 6 contain brief discussions of the efficiency of REE and the existence of partially revealing REE, respectively. 2

Rational Expectations Equilibrium

There are n traders, indexed by the superscript /, and € commodities, indexed by the subscript j . Each trader / has a set Sl of private information signals, where Sl is a finite-dimensional Euclidean space. Let S = 11*1=1 S\ with generic element s = (s1, . . . , sn). An element s e S will be termed a state of information. Let n denote a Borel probability measure on 5. Let U denote the set of smooth (infinitely continuously differentiable) functions u: R{+ —> R satisfying, for each x e Ri+, (i) Dul(x)eRi+, 2 (ii) D u!(x) is negative definite, and (iii) the closure in Re of the upper contour set {xf e Ri+: u(x') > u(x)} is contained in Re++. The set U is topologized as a subspace of C°(Rl+,R) with the C00 compact-open topology.1 Each trader / has a state-dependent utility func1

This is the topology of C00 uniform convergence on compact subsets of Ri+, which requires that similar utility functions be uniformly close and have uniformly close mixed partial derivatives of all orders on any closed and bounded subset of strictly positive commodity bundles xe Ri+. However, for all of the results described in this paper except those based on Allen (1982b) and Jordan (1982b), we could just as well have used utilities that are required only to be twice continuously differentiable; in that case, the subset

Rational Expectations Equilibrium: A Retrospective

45

tion vl: S^>U that is Borel-measurable. The tactic of representing the utility function as a random variable, as opposed to viewing the utility of each commodity bundle as a separate random variable, is taken from Allen (1981a). It avoids a technical problem in the definition of conditional expected utility maximization noted by Kreps (1977).2 In general, each trader f s endowment can also be state-dependent (but measurable with respect to /)•However, in this exposition, we will adopt the common simplifying assumption that each trader / has a state-independent endowment ele R{+. Let A denote the unit simplex in R{, and let M = [(p^y1, . . . , / ) e A X RnL. 2?=i y = 0 and pyl = 0 for each /} denote the set of prices and n-tuples of net trades that satisfy the aggregate resource balance and individual budget constraints. A rational expectations equilibrium will take the form of a Borel-measurable function from S to M. The space M contains the market signals that traders can use to augment the information provided by their private signals. In a rational expectations equilibrium, the information available to trader / will be at least sl and at most the entire state s of information. It will be useful to define equilibria corresponding to the two extremes. A private information equilibrium (PIE) is a Borel measurable function (/?(•), (y'(-))/): S^M satisfying, for each /, maximizes E\V1 subject to p(s\yl = 0 for a.e. s,

(1)

l

where E{V I s ) e U is the conditional expectation of V given the a-field generated by the projection s—>s l. A full communication equilibrium (FCE) is a Borel-measurable function (/?(•),(j?('))()- S-^M satisfying, for each /, yl(s) maximizes v'Mfe1 + yl) subject to p(s\yl = 0 for a.e. s. 2

2

2

(2) 2

U of C (R{+, R) would be given the (C compact-open) topology of C uniform convergence on compacta. Similarly, we could use the more complicated C00 Whitney topology, which imposes very strong similarity requirements "at infinity" and as we approach the boundary of the positive orthant in Re for functions and all of their derivatives. Roughly, the problem is that there are both a continuum of commodity bundles and a continuum of price values on which to condition expected utility. If the utility of each commodity bundle is viewed as a separate random variable, then - since the conditional expected utility of each bundle is arbitrary on a set of price values having probability measure 0 - for each price some bundle can be given an arbitrarily high conditional expected utility, making nonsense of conditional expected utility maximization. Kreps (1977) suggests denning conditional expected utility as an integral with respect to a fixed regular conditional probability distribution in order to avoid this problem, but viewing the utility function as a single random variable works just as well. For further details, see the appendix on measurability of demand in Allen (1981a).

46

Beth Allen and James S. Jordan

In the conventional definition of rational expectations equilibrium, traders augment their private signals with the information revealed by market prices. Formally, a rational expectations equilibrium (REE) is a Borel-measurable function (p(-), (yf'(-))/): S-^M satisfying, for each /, yl(s\ maximizesE\v l sl,pis)\[el + yl) subject to p[s\yl = 0 for a.e. s. 3

(3)

Existence and Nonexistence of REE

The earliest simple example of the nonexistence of REE was given by Kreps (1977). In this example, there are only two possible states. Since there are only two states, if prices differ between the states then an REE must be an FCE, whereas if prices are the same in the two states then an REE must be a PIE. Nonexistence of REE is proved by showing that in the unique FCE, the prices are the same in the two states, and in the unique PIE, the prices differ between the two states. Formally, consider a variant of the counterexample that appears in Allen (1984). Suppose there are two traders and two commodities, and suppose that the support of the probability measure n on traders' signals consists of the two-point set {(sla,s2), (slb, s2)}, where s\±s\» and that the two states have equal probability. Thus trader 1 is privately fully informed of the state, while the private signal of trader 2 does not reveal the state. Let sa = (sl,s2) and sb = (si,s2), and define v1 and v2 on these states as follows: =21njci+lnjt 2, =lnjti + 21njc2, =lnx! +21nx 2 , The state-independent endowments are e1 = e2 = (3,3). Then the unique FCE satisfies p(sa) = p(sb) = {111, 1/2) while y\sa) = (1, - 1 ) , y\sb) = ( - 1 , 1 ) , and y2(-) = —y 1^). Since p(sa) =p(sb), the FCE cannot be an REE. In a PIE, trader 2 is uninformed of the state and so, for any given price, trader 2 will have the same demand in both states. It is also clear (from the definition of v1 and the fact that trader 1 is privately informed of the state) that, for any given price, trader 1 will demand more of commodity 1 in state sa than in state sb. Thus no PIE can have the same prices in both states; in fact, the unique PIE has p(sa) = (7/12,5/12) and p(sb) =

Rational Expectations Equilibrium: A Retrospective

47

(5/12,7/12). Hence there is no rational expectations equilibrium for this economy. Radner (1979) also developed an example with log-linear utility functions. However, in his framework, the information signals concern traders' conditional probabilities of various events, while the relationships between states of the world and agents' preferences are considered fixed. We find it simpler to identify signals and states of the world, thereby replacing Radner's conditional probabilities with Os and Is. As in Allen (1981a, 1984), we consider a space of economies that is defined by state-dependent preferences. By contrast, Radner's (1979) model fixes the state-dependent preferences and then specifies various economies by their collections of conditional probabilities. Radner (1979) showed that the property that prevents an FCE from being an REE in such examples is nongeneric in the space of arrays of conditional probabilities of states of the world, given agents' signals. Radner's result was stated in the context of financial asset markets, but the essential idea is easily expressed in the present model. For each state of information s, the economy (e\vl(s)(-))%i is a smooth economy as defined by Debreu (1972). Suppose that the support of n is a finite set S° = [Sk)k=i- Then an FCE consists simply of a Walrasian equilibrium (p(sk), (yl(sk))i) for the economy (el,vl(sk)('))» for each k. If p(-) is oneto-one then the FCE is an REE as well. The contrary condition that p{sk) = p(sk) for some k and k' is a "knife-edge" property that Radner showed is nongeneric in the smooth economies {(e\ v/(^)(-))/}f=i parameterized as described here. A formal statement and proof of the analog of this result for the present model is given by Allen (1984). Allen (1981a and 1982b) showed that the generic existence of fully revealing rational expectations equilibria extends well beyond the class of economies considered by Radner (1979). Whereas Radner's analysis relied heavily on the finiteness of the set of possible states of information, Allen showed that the essential condition is merely that the price have a higher dimension than the state. Allen (1981a) established that if dim S < (1/2) dim A, then an FCE price function /?(•) is generically oneto-one. Moreover, Allen (1982b) showed that if dim S < dim A, then an FCE price function is generically one-to-one on the complement of a subset of S having Lebesgue measure 0. In particular, if n is absolutely continuous with respect to Lebesgue measure, then an FCE price function is generically one-to-one on a set having probability 1 and is thus generically an REE. The first result relied on the Whitney embedding theorem from differential topology while the second used the related ideas of self-intersections of immersed manifolds (based on multijet trans versality).

48

Beth Allen and James S. Jordan

Stimulated by Allen's results, Jordan and Radner (1979,1982) considered the equal-dimension case, that is, dim S = dim A. They showed via an example that in this case neither the existence nor the nonexistence of an REE is a generic property. This left the higher-dimensional case, dim S > dim A. In this case, an FCE generically cannot be an REE, since an FCE price function/? generically is smooth enough that it cannot map S to a lower-dimensional space and still be one-to-one on a subset of S with full Lebesgue measure. However, Jordan (1982b) demonstrated the generic existence of REE price functions p that are two-to-one. Moreover, the price functions p can be made arbitrarily close to one-toone in the sense that the pairs of states mapped to the same price can be made uniformly arbitrarily close to one another. In particular, there are many such equilibria. However, the constructed price functions are discontinuous on a dense subset of 5, so one would hesitate to interpret such equilibria as the natural outcome of a market mechanism. Thus, in the higher-dimensional case, the generic existence of economically plausible rational expectations equilibria remains an open question.

4

Generalized Rational Expectations Equilibrium

In the conventional definition of rational expectations equilibrium, prices are the only market variables that traders use to augment their private information. This accords with the usual interpretation of prices as the coordinating signals that enable the competitive market to achieve Pareto optimal allocations. Indeed, Grossman (1981) argues that the existence of a fully revealing REE is the appropriate formalization of Hayek's famous assertion that the market mechanism is informationally efficient (Hayek 1945). This interpretation suggests that the informational discontinuity discovered by Radner constitutes a serious flaw in the market mechanism that may preclude the possibility of informationally efficient price systems. However, there is a prior line of research on Hayek's conjecture that suggests that the sole reliance on price constitutes a misunderstanding of the market mechanism. Hurwicz (1977) and Mount and Reiter (1974) have constructed an explicit model of competitive equilibrium as an informationally decentralized allocation mechanism. In their model, the market signal, which these authors call the "competitive message," is the entire list of market variables, (p, (/);). Moreover, they show that it is not possible for a decentralized allocation mechanism to use a message of smaller dimension, such as the price vector alone, and still achieve Pareto optimal allocations for all classical pure exchange environments. The Hurwicz (1977)

Rational Expectations Equilibrium: A Retrospective

49

and Mount and Reiter (1974) model is nonstochastic, but the reasoning behind the result can be illustrated using the two-person, twocommodity example of the previous section. Let M be an arbitrary set of equilibrium messages, let Y = { ( / , y2) e R2 X R2 \yl + / = 0}, and let h: M —> Y be an arbitrary outcome function. Fix the endowments at el = e2 = (3,3) as before. A (nonstochastic) exchange environment is then specified by a pair of utility functions (w1, u2) e Lf2. An allocation mechanism consists of a correspondence pt: {T2-^—>M, called the equilibrium message correspondence, together with the outcome function h:M ->Y. The correspondence ju associates with each (w1, U2) a set //(w1, u2) of equilibrium messages, and the function h associates with each m e ^(u1, u2) a net trade outcome (yx,y2) = h(m). In order to represent the competitive mechanism in this way, we require h(ju(ul, u2)) to be the set of competitive equilibrium trades for each (ul,u2). Obviously, each equilibrium message m e ju(u\ u2) must contain enough information about (u\ u2) to determine the competitive equilibrium net trade. In fact, decentralization requires that messages contain still more information. At this level of abstraction, decentralization means simply that the concept of equilibrium can be decomposed into separate equilibrium conditions for each trader. Formally, the allocation mechanism (ju,M,h) is informationally decentralized if there are individual equilibrium message correspondences JU1,^2: £/—>—> M such that, for every (ul,u2), (D) For example, to define the competitive mechanism (pic,Mc,hc), let Mc = M, let hc be the projection hc: (p,yl,y2) —> (yl,y2), and define julc(ul) to be the set of market signals (p',yl,y2) such that yl maximizes u(el + z) subject to pz = 0. Then use (D) to define juc by setting juc(ul, u2) = ptlc(ul) Djul(u2). In effect, julc(ul) is trader fs offer curve, and the set of competitive equilibrium messages is defined as the intersection of the offer curves. The decentralization condition (D), as a property of/i, is existential, since it requires the existence of suitable correspondences ju\ However, it is easily seen that the following property is necessary and sufficient for the existence of correspondences JU1 and JU2 satisfying (D): /*(wi, u2) H ju(ul ul) =

IJL[U\M) 2

H ju(ul,u2a) 2

(D')

for every pair of environments (u\, u a), (ul, u b) (Mount and Reiter 1974, Lemma 5). To see why decentralization requires the equilibrium messages to contain more information than just the equilibrium trades, consider the utility functions used in Section 3:

50

Beth Allen and James S. Jordan u\(x\ =

uUx) = u2(x) = lnx! +21nx 2 , ub [x\ = 2 In Xi + In x2. Note that in the environments (wj, «4) and (w£, ul) there is no trade in equilibrium. Therefore, if the equilibrium messages were just the equilibrium trades, we would have fi(u\,u2b) np(ul,ul) = {(0,0)}, which contradicts (D') because the environments (ul, u2a) and (ul, u2b) have nonzero equilibrium trades. The reason for this digression into informational decentralization theory is the rather surprising result that a version of (D') is necessary for the existence of rational expectations equilibrium. Consider the stochastic exchange environment described in Section 3, but suppose now that the expectations of the uninformed trader are conditioned on the equilibrium message of some allocation mechanism (ju,M,h). To ensure that conditional expectations are well-defined, suppose that the equilibrium message correspondence ju is a single-valued function. Then a fully revealing equilibrium exists if pi(ul,ul) J=ju(ub,u2b) (recall that v 1 ^ ) = u\, v2(sa) = u2a, etc.), because in that case the equilibrium message reveals the state to trader 2. However, if, as is the case with the equilibrium price, ti(u\, ul) = ju(ul, u2b), then only a nonrevealing equilibrium is possible, the existence of which requires that ju(ul, nau2a + (1 - 7ra)u2b) = pt(ub, nau2a + (1 - 7ia)u2b), where na denotes the probability of the state sa. Hence the existence of a "message-conditional expectations equilibrium" requires that

=> ju(ula, na u2 + (l - na )u2b) = n{u\, nau2 +(l-Ka

)u2b j .

Suppose we further require that equilibrium exist for all values of na. If we assume that the function pi is continuous with respect to na and let Ka —> 1, we obtain the implication fi{lila, U2} = fl{u\, Ub}=$ fi{lila, U2^ = li{u

Letting na —> 0 strengthens the implication to \, U2a ) =

If we further require that equilibrium exist for the stochastic environments obtained by permuting the state-dependent utilities to v2(sa) = u\ and v2(sb) = ul, then we add the reverse implication to obtain

Rational Expectations Equilibrium: A Retrospective

51

fi(u\, ul) = ju(ulb, ul) « / ^ , w2) = ju(u\, w2) = /*(ai, ^ 2 ), which is simply (D') for single-valued ju. Thus the decentralization condition (D') is necessary for the existence of a message-conditional expectations equilibrium for every two-state stochastic environment, provided that /u is single-valued and continuous in na. The assumption that pi is single-valued throughout its domain U2 is much too restrictive. For the purpose of establishing necessary conditions for the existence of equilibrium, however, we can confine the domain of pt to the set of log-linear utility pairs. On this domain, the competitive mechanism fic: (ul,u2) —> (p,(yl)i) is single-valued and also has the required continuity with respect to the utility coefficients. Moreover, any JU obtained as a continuous function of the competitive message also inherits these properties. That is, if the market variables (p, (yl){) are condensed via any continuous function /: MC^>M, the resulting JU(-) = f(juc(-)) must satisfy (D') on log-linear utility pairs as a necessary condition for the general existence of equilibrium with expectations conditioned on the condensed market data. In particular, an REE fails to exist in the example of Section 3 because condensing the competitive message (/?, (/),) to the price alone violates (D'). It is natural to ask whether other condensations might be consistent with (D'). Unfortunately, Jordan (1977, Thm. 5.4) shows that there is no other continuous condensation, except for the trivial case of the constant function, that satisfies (D'). However, as will be shown in what follows, the force of this negative result can be lessened by permitting different traders to condition their expectations on different condensations. Fortunately, the contribution of informational decentralization theory to rational expectations equilibrium is not limited to nonexistence results. It also suggests that, in order to obtain general existence theorems, we should extend the concept of rational expectations equilibrium to allow traders to condition their expectations on additional market variables. More formally, suppose that each trader / has a set of market signals M that are generated by a function/': M—>M'. Define a generalized rational expectations equilibrium (GREE) as a Borel measurable function (/?(•), Cv'(-))/): S^M satisfying, for each /,

yl(s) maximizes E\V1 S\ f{[p(s\ (y'W) J)}(*' + subject to p(s)yl = 0 for a.e. s.

(4)

Taking each f to be the projection (p, (/'),-) -^p reduces (4) to (3). Alternatively, suppose that each f is the identity, so that each trader

52

Beth Allen and James S. Jordan

observes the entire competitive message. Then one can show that an FCE is always a GREE. In fact, the same is true if each f is taken to be the projection (p, (y);) -> (p,/). The proof of this is nearly immediate. Suppose that (p(*)> (9l('))d *s a n FCE. We need to show that, for each /, yHs) maximizes EW S\ p(s\ yHs^iie1 + / ) subject to p(s)yl = 0 for a.e. s.

(5)

In other words, we need to show that when trader /'s information is reduced from s to (sl,p(s),yl(s)), the trade yl(s) remains expected utility maximizing. Given any (s,/?, y'), let S = {s I sl = T, p(s) = p, and yl(s) = y1}. Thus S is the set of states that trader / cannot distinguish from one another. The FCE condition (1) states that, for almost every s e 5, yl maximizes v\s){^ + /') subject to pyl = 0. It follows immediately that yl maximizes any convex combination of the functions {/ —»vl(s){el + yl): s e S} subject to pyl = 0. The conditional expectation in (5) is simply one such convex combination. Thus the market data (p,yl) are informationally sufficient to support an FCE as a GREE. It may be interesting to note that this proof is very similar to the usual demonstration that a correlated equilibrium of a game can be supported by using the players' actions as the correlating messages (Aumann 1987). The proof indicates that the conditional expectation in (5) can be conditioned on the market data (^(s1),)^)) alone, so that the private signal sl is unnecessary. However, it is not possible to reduce the data to yl(s) alone, since a trade that is optimal in different states at different prices may fail to maximize the trade-conditional expected utility at any of the respective prices. It is essential to the foregoing proof that the budget constraint pyl = 0 be the same for all s G S. In fact, Jordan (1982a, Thm. 2.5) shows that the data (/?,/) cannot be further condensed, except to a constant, without losing the general existence of equilibrium. The interpretation of (5) seems troublesome. How can a trader choose a trade to maximize expected utility conditional on the trade itself? This question can be blunted somewhat by adopting -2yV:/y as a euphemism for yl or by recalling that, in auction theory, bidders are assumed to avoid the "winner's curse" by choosing a bid that maximizes their expected payoff conditional on the bid being accepted. However, Beja (1976) has shown that when one attempts to formalize an REE as a Bayesian Nash equilibrium, as in auction theory, in which traders choose excess demand functions, the existence problem is severe. Thus, although the equilibrium concept defined by (5) completely neutralizes the informational discon-

Rational Expectations Equilibrium: A Retrospective

53

tinuity, it is not persuasive as a self-contained description of the process by which markets aggregate and communicate information. One approach to interpreting (5) is to model more explicitly the dynamics of information transmission. A particularly simple model, first suggested by Reiter (1976), is the following. Suppose that traders initially use only their private information, so that the market initially moves to a PIE. However, before the PIE trades are executed, traders add to their private information the information revealed by the PIE prices. This will typically change traders' expectations, requiring a new round of trade, and so on. In the example described in Section 2, the PIE prices p(sa) = (7/12,5/12) and p(sb) = (5/12,7/12) reveal the state to the uninformed trader, so the second and final iteration is the FCE. When there are more than two possible states, more than two iterations will typically be required to reach a price function p{-) that reveals no further information. In general, the final iteration need not be an FCE. The two-state example of Section 2 has the interesting property that the final price function, in this case the FCEp(sa) = p(sb) = (1/2,1/2), does not itself reveal the information revealed by earlier prices. Instead, the information is revealed by the trade function y2{-). Jordan (1982c, Thm. 5.7.C) shows that, in general, the information revealed by earlier prices that is decision-relevant to trader i is revealed by the pair (/?(*)>y'O) a t the final iteration. Thus one can interpret (5) by viewing the pair (p,yl(-)) as a representation of the decision-relevant information revealed to trader / during the trading process. A very similar trading process was formulated independently by Kobayashi (1977) in a financial asset market model under the assumptions that (i) traders have constant absolute risk aversion and (ii) the joint distribution of future asset values and traders' private information is normal. These assumptions guarantee the existence of an R E E (Grossman 1976 and 1978), and Kobayashi proved that the trading process reaches the REE in a number of iterations equal to the number of traders. In Kobayashi's model, the trades at each iteration are actually executed, but the assumption of constant absolute risk aversion prevents the resulting capital gains or losses from having any effect on the process. In both models, it is critical that traders behave myopically and choose their demand as though each iteration were the last. If traders attempt to anticipate future prices, the informational discontinuity reappears, albeit in a more subtle form (see, e.g., Border and Jordan 1979, Dubey, Geanakoplos, and Shubik 1987, and Futia 1981). For example, consider a finite-period stochastic exchange environment, with only spot markets in each period, in which traders' preferences are not

54

Beth Allen and James S. Jordan

intertemporally separable. Because future consumption affects preferences for current consumption, a trader will attempt to anticipate future prices when choosing current demand. Suppose that traders condition their expectations on private information and past prices. Then current prices influence future information, and thus influence future prices. Traders' rational expectations of future prices, conditional on current information, influence current demands, which determine current prices. Because of this feedback, the discontinuous relation between prices and information can again prevent the existence of equilibrium. 5

Efficiency Properties of REE

One reason for dissatisfaction with fully revealing rational expectations equilibria and the microeconomic models that generate them is that, realistically, such equilibria are too good to be true. The situation described in Sections 2 and 3 contradicts our economic intuition in a significant way. We do not believe that prices can transmit all information of interest to economic agents because information may be extremely complicated, with different types of information being most relevant for different trades. No finite number of (even continuously variable) prices for commodities that are transacted in markets can completely capture this complex information. If all useful information were to be carried by equilibrium prices, the equilibrium would be at least informationally efficient and, in the absence of other potential market distortions, one could perhaps expect the resulting allocation to be optimal in a wider sense. However, full Pareto optimality cannot be expected if traders wish to share their risks - insurance markets may be desirable yet precluded by information. Hence, we must restrict efficiency properties by stressing that we only consider efficiency relative to existing markets as well as with respect to the given information structure. With these two important qualifications, fully revealing rational expectations equilibrium allocations are necessarily ex ante Pareto optimal, as observed by Radner (1967). The situation is rendered more complicated by an important point elucidated, by means of an example, in Green (1981). In a general equilibrium model, more information may be undesirable unless complete information is achieved; additional partial information can give rise to Pareto inferior equilibrium allocations because the new information can destroy some previously feasible opportunities for risk sharing. Essentially, the presence of information can destroy insurance markets, thereby producing inefficiency unless all agents can freely obtain complete information.

Rational Expectations Equilibrium: A Retrospective

55

A further examination of the optimality of rational expectations equilibrium allocations appears in Laffont (1985). This article considers the issue of whether partially revealing rational expectations equilibrium allocations satisfy constrained efficiency. Unfortunately, the answer is generally negative. In partially revealing cases, competitive markets do not lead to the transmission of those parts of consumers' information that matter most for achieving Pareto improving trades. Even if equilibrium prices transmit the maximal amount - in a dimensional sense - of information that can be communicated given the number of commodities present in the economy, the information that is carried by prices could be such that it doesn't affect traders' utilities very much. In short, one cannot expect the most important information to be conveyed by prices if prices cannot transmit all information. The "efficient markets hypothesis" article of Jordan (1983) makes a similar point. Only for three special parametric classes of utility functions - linear, logarithmic, and exponential - do we find informationally efficient rational expectations equilibria. Moreover, these economies obviously constitute "knife-edge" examples; they are not generic in any reasonable sense. 6

Partially Revealing REE

In addition to the objection discussed in the previous section - that informationally efficient rational expectations equilibria cannot be expected to arise in most economic situations - fully revealing R E E are inconsistent in a more inclusive economic model in which the information is endogenous. Specifically, Beja (1976) argues that strategic players in a REE market game would not gather information and, as a result, information cannot be conveyed by prices. A similar point was made by Grossman and Stiglitz (1976) in a parametric model of a market for a financial asset. They observe that if information can be freely communicated by prices then no trader would ever pay to gain information, and so REE models with endogenous costly acquisition of information are therefore not consistent. The standard model of partially revealing equilibrium in financial asset markets is that of Hellwig (1980), which generalizes and extends the model of Grossman and Stiglitz (1976, 1980). Traders are assumed to have constant absolute risk aversion, and the joint distribution of private information and future asset values is assumed to be normal. The full revelation of private information is prevented by the device of adding a random noise term to the aggregate supply, so that price movements can be caused by either private information or supply shocks. Unfortunately,

56

Beth Allen and James S. Jordan

the existence of the partially revealing REE appears to depend on the special assumptions of this model. The fact that the original nonexistence example of Green (1977) was created by adding noise to traders' endowments suggests that there is little hope that endowment noise could serve as the basis for a general theory of partially revealing REE. Progress in analyzing the possibilities for genuinely partially revealing rational expectations equilibria has been disappointing in general. An explicit example of partially revealing rational expectations equilibrium prices was provided in Allen (1981b), but that class of economies suffers from the observation that the combination of price information and any agent's initial private information corresponds to full information. Thus, the rational expectations equilibrium allocations exactly equal the full communication demands evaluated at the resulting market prices. More recently, Ausubel (1990) has studied partially revealing rational expectations equilibrium, albeit for a somewhat special economic model. More generally, results for the partially revealing case all feature some form of approximation, either to exact market clearing or to complete rationality in agents' use of information. To see that slight deviations from precise market clearing greatly simplify the situation, consider our basic model when there are many more (in a dimensional sense) parameters than prices and when (for simplicity) all agents are either fully informed or completely uninformed. Divide the compact parameter set into sets of small diameter and choose a single distinct price for each set in the partition so as to make demand arbitrarily close to zero when the uninformed agents condition on the partition. This forms a rational expectations approximate equilibrium which, by definition, is not fully revealing. A similar trick can be used to construct rational expectations approximate equilibria based on those in Allen (1981a, 1982b) rather than the finite-state case of Radner (1979) or Allen (1984). See Allen (1982a) for details. A more complicated approach based on the general idea of "noise" or dispersion was pursued in simultaneous and independent work by Allen (1981c, published as 1985a,b) and by Anderson and Sonnenschein (1981, published in the 1982 Journal of Economic Theory symposium issue). One interpretation of this research is that it illustrates the tradeoff between exact rationality and exact market clearing in the partially revealing context with noise. These papers take a different approach to the addition of noise than Hell wig (1980), who adds noise to supplies whereas Allen (1981c, 1985a,b) and Anderson and Sonnenschein (1981, 1982) use the concept of noise to alter the definitions of rational expectations equilibria in various ways - that is, noisy rationality or noisy

Rational Expectations Equilibrium: A Retrospective

57

market clearing. Unfortunately, this literature provides only approximations to rational expectations equilibrium. The big question of formulating a satisfactory and consistent concept of partially revealing rational expectations equilibrium has remained open. An additional desideratum is that these equilibria display the usual properties of the equilibrium price correspondence and, ideally, the same decentralization or implementation features as in the standard model of an exchange economy without asymmetric information. 7

Conclusion

Radner's influence on the general equilibrium theory of rational expectations was seminal. Radner (1967,1968,1972) provided the first models of trade among agents with differing private information, the first analysis of the revelation of private information by equilibrium prices, and the first discovery of the informational discontinuity that creates the equilibrium existence problem. Radner (1979) established the first result on the generic existence of equilibrium. Now that subsequent researchers on the existence problem appear to have followed Radner's example of moving on to other topics, it is appropriate to take stock of the results (see Allen 1986 and Radner 1982). The existence, if not the economic interpretation, of fully revealing equilibrium is well established. In the case of complete markets, Grossman (1981, Thm. 2) shows the general existence of fully revealing equilibrium. In the usual incomplete markets model, other results mentioned here establish the generic existence of fully revealing equilibrium in the lowerdimensional case and of nearly fully revealing equilibrium in the higherdimensional case. More generally, expanding the market signal to include net trades, as in (5), will always produce a fully revealing equilibrium whenever a full communication equilibrium exists, which requires only the classical assumptions. The existence of partially revealing equilibrium, in contrast, remains precarious. Results mentioned in Section 6 show that, in various senses of approximation, there exist approximate equilibria with intuitively reasonable revelation properties. With respect to exact equilibria, however, the authors are aware of no existence results that do not rely on very special assumptions on preferences (e.g., constant absolute risk aversion) or the stochastic structure of information (e.g., joint normality). When one reflects on the pervasive use of partially revealing equilibrium in the theory of financial asset markets (see, e.g., Black 1986), the absence of a firm foundation for partially revealing equilibrium is especially disquieting.

58

Beth Allen and James S. Jordan References

Allen, Beth, 1981a, Generic existence of completely revealing equilibria for economies with uncertainty when prices convey information, Econometrica 49,1173-99. 1981b, A class of monotone economies in which rational expectations equilibria exist but prices do not reveal all information, Economics Letters 7, 227-32. 1981c, The existence of expectations equilibria in a large economy with noisy price observations, Working Paper no. 81-01, Center for Analytic Research in Economics and the Social Sciences, Department of Economics, University of Pennsylvania, Philadelphia. 1982a, Approximate equilibria in microeconomic rational expectations models, Journal of Economic Theory 26, 244-60. 1982b, Strict rational expectations equilibria with diffuseness, Journal of Economic Theory 27, 20-46. 1984, Equilibria in which prices convey information: The finite case, M. Boyer and R. Kihlstrom (eds.), Bayesian Models in Economic Theory. Amsterdam: North Holland, pp. 63-92. 1985a, The existence of rational expectations equilibria in a large economy with noisy price observations, Journal of Mathematical Economics 14, 67-103. 1985b, The existence of fully rational expectations approximate equilibria with noisy price observations, Journal of Economic Theory 37, 213-53. 1986, General equilibrium with rational expectations, W. Hildenbrand and A. Mas-Colell (eds.), Contributions to Mathematical Economics in Honor of Gerard Debreu. Amsterdam: North-Holland, pp. 1-23. Anderson, Robert M., and Hugo Sonnenschein, 1981, On the existence of rational expectations equilibria, Mimeo, Department of Economics, Princeton University, Princeton, NJ. 1982, On the existence of rational expectations equilibrium, Journal of Economic Theory 26, 261-78. Aumann, Robert, 1987, Correlated equilibrium as an extension of Bayesian rationality, Econometrica 55,1-18. Ausubel, Lawrence M , 1990, Partially-revealing rational expectations equilibrium in a competitive economy, Journal of Economic Theory 50, 93-126. Beja, Abram, 1976, The limited information efficiency of market processes, Working Paper no. 43, Research Program in Finance, University of California, Berkeley. Black, Fisher, 1986, Noise, Journal of Finance 41, 529-43. Border, Kim, and J. S. Jordan, 1979, Expectations equilibrium with expectations conditioned on past data, Journal of Economic Theory 22, 395-406. Debreu, Gerard, 1972, Smooth preferences, Econometrica 40, 603-15. Dubey, Pradeep, John Geanakoplos, and Martin Shubik, 1987, The revelation of information in strategic market games: A critique of rational expectations, Journal of Mathematical Economics 16,105-37.

Rational Expectations Equilibrium: A Retrospective

59

Futia, Carl A., 1981, Rational expectations in stationary linear models, Econometrica 49,171-92. Green, Jerry R., 1973, Information, efficiency and equilibrium, Discussion Paper no. 284, Harvard Institute of Economic Research, Harvard University, Cambridge, MA. 1975, The non-existence of informational equilibria, Discussion Paper no. 410, Harvard Institute of Economic Research, Harvard University, Cambridge, MA. 1977, The non-existence of informational equilibria, Review of Economic Studies 44, 451-63. 1981, Value of information with sequential futures markets, Econometrica 49, 335-58. Grossman, Sanford J., 1976, On the efficiency of competitive stock markets where trades have diverse information, Journal of Finance 31, 573-85. 1978, Further results on the informational efficiency of competitive stock markets, Journal of Economic Theory 18, 81-101. 1981, An introduction to the theory of rational expectations under asymmetric information, Review of Economic Studies 48, 541-59. Grossman, Sanford J., and Joseph E. Stiglitz, 1976, Information and competitive price systems, American Economic Review Papers and Proceedings 66, 246-53. 1980, On the impossibility of informationally efficient markets, American Economic Review 70, 393-408. Grunberg, E., and F. Modigliani, 1954, The predictability of social events, Journal of Political Economy 62, 465-78. Handel, Christopher J., 1980, Allocational efficiency in a competitive capital market, Ph.D. Dissertation, Stanford University. Hayek, F, 1945, The use of knowledge in society, American Economic Review 35, 519-30. Hellwig, Martin F, 1980, On the aggregation of information in competitive markets, Journal of Economic Theory 22, 477-98. Hurwicz, L., 1977, On the dimensional requirements of informationally decentralized Pareto-satisfactory processes, K. J. Arrow and L. Hurwicz (eds.), Studies in Resource Allocation Process. Cambridge University Press. Jordan, J. S., 1977, Expectations equilibrium and informational efficiency in stochastic environments, Journal of Economic Theory 16, 354-72. 1982a, Admissible market data structures: A complete characterization, Journal of Economic Theory 28,19-31. 1982b, The generic existence of rational expectations equilibrium in the higher dimensional case, Journal of Economic Theory 26, 224-43. 1982c, A dynamic model of expectations equilibrium, Journal of Economic Theory 28, 235-54. 1983, On the efficient markets hypothesis, Econometrica 51,1325-44. Jordan, J., and R. Radner, 1979, The nonexistence of rational expectations equilibrium: A robust example, Mimeo, Department of Economics, University of Minnesota, Minneapolis.

60

Beth Allen and James S. Jordan

1982, Rational expectations in microeconomic models: An overview, Journal of Economic Theory 26, 201-23. Kobayashi,T, 1977, A convergence theorem on rational expectations equilibrium with price information, Working Paper no. 79, The Economic Series, Institute for Mathematical Studies in the Social Sciences, Stanford University. Kreps, David M., 1977, A note on "fulfilled expectations" equilibria, Journal of Economic Theory 14, 32-43. Laffont, J. J, 1985, On the welfare analysis of rational expectations equilibrium with asymmetric information, Econometrica 53,1-29. Mount, K., and S. Reiter, 1974, The informational size of message spaces, Journal of Economic Theory 8,161-92. Muth, John R, 1961, Rational expectations and the theory of price movements, Econometrica 29, 315-35. Radner, Roy, 1967, Equilibre des marches a terme et au comptant en cas d'incertitude, Cahiers d'Econometrie (C.N.R.S., Paris) 4, 35-42. [Translated as: Equilibrium of spot and futures markets under uncertainty, Technical Report no. 24, Center for Research in Management Science, University of California, Berkeley.] 1968, Competitive equilibrium under uncertainty, Econometrica 36, 31-58. 1972, Existence of equilibrium of plans, prices, and price expectations in a sequence of markets, Econometrica 40, 289-303. 1979, Rational expectations equilibrium: Generic existence and the information revealed by prices, Econometrica 47, 665-78. 1982, Equilibrium under uncertainty, K. J. Arrow and M. D. Intriligator (eds.), The Handbook of Mathematical Economics, vol. II. Amsterdam: NorthHolland, pp. 923-1006. Reiter, Stanley, 1976, On expectations equilibrium, Mimeo, Northwestern University, Evanston, IL.

CHAPTER 3

Rational Expectations and Rational Learning Lawrence E. Blume and David Easley

1

Introduction

The issue of expectation formation arises naturally in economies with a sequence of spot and incomplete futures markets. In such economies, individuals use forecasts of future prices in order to make decisions about current consumption and investment. Equilibria in the markets for current consumption and currently available futures contracts may exist even if individuals have differing and incorrect expectations. Such equilibria are called temporary equilibria, and other than the usual conditions on preferences and endowments, their existence requires only weak conditions on expectations about future prices. However, absent any structure on price expectations, there is little more to be said about equilibria at any date. One attempt to tie down expectations is the rational expectations hypothesis: that individuals hold common, correct price expectations in economies with incomplete markets. The existence of rational expectations equilibria was first addressed by Roy Radner in his seminal 1972 paper on plans, prices, and price expectations. Radner fixed expectations by requiring that agents hold common price expectations and that their plans be consistent with market clearing at all future dates at prices equal to their expected prices. The term "self-fulfilling" has been used to describe these "rational" expectations. It is particularly apt because it emphasizes that the actual sequence of prices is determined by the expectations agents use. In the 1972 model, payoff-relevant information was distributed We thank participants at the June 1992 Economic Theory Workshop in honor of Roy Radner, Cornell University, and two anonymous referees for their helpful comments. We also want to thank seminar participants at CORE, Erasmus University, Tel Aviv University, and The Technion for helpful discussions. Finally, we thank Mukul Majumdar for organizing the Radner workshop and for his help in constructing this article. Financial support from NSF Grant SES-8921415 is gratefully acknowledged.

61

62

Lawrence E. Blume and David Easley

"symmetrically"; that is, all traders had the same information. Consequently there is nothing to be learned from other traders, and so in equilibrium there is nothing to be inferred from current prices about prices in the future. In 1979, Radner demonstrated the existence of a rational expectations equilibrium when payoff-relevant information is "asymmetrically" distributed - that is, when not all players have the same information. In these equilibria, traders infer the information of others by observing current-period prices. It is inconceivable that individuals should be born with the complete understanding of the economy required by rational expectations equilibria in either model. Thus it is natural to ask if traders can learn to hold rational expectations. Outside of a rational expectations equilibrium, traders entertain different models of the relationship between present prices and private information. They use equilibrium prices to help identify the correct models. Blume and Easley (1982) was the first paper to address this learning problem in a general equilibrium setting. The answer offered to the learning question in that paper is "no" for a large class of learning models. Rational expectations equilibria are locally but not globally stable, under reasonable learning dynamics, and other non-self-fulfilling equilibria may also be locally stable. Convergence fails because the likelihood functions used to assess the relative worth of various models are correctly specified at equilibria but are not correctly specified far away from equilibrium. On the other hand, Blume and Easley (1984) show that, when the likelihood function is correctly specified away from equilibria as well as near, Bayesian estimation of the true models is consistent; that is, Bayesian traders learn rational expectations and prices converge to those of a rational expectations equilibrium. The origin of the likelihood function raises an important problem that has been ignored until recently. In order to carry out the Bayesian analysis in Blume and Easley (1984), the likelihood function for each trader must describe the correct likelihood of various paths for the evolution of prices. This likelihood depends upon a number of parameters, including the prior beliefs and likelihood functions of the other traders. Thus the mutual consistency of correctly specified likelihood functions and prior beliefs is a kind of equilibrium existence problem. In Section 4 we define appropriate equilibrium stochastic processes and call them Bayesian equilibrium processes (BEPs). (A similar construct arises in noncooperative game theory: in Section 3, we define and discuss Bayesian strategy revision processes.) The existence question is easy to solve. Rational expectations equilibria are BEPs, but these are trivial equilibria in which no learning about models takes place. The important

Rational Expectations and Rational Learning

63

question to ask, which we duck in this paper, is how rich is the set of prior beliefs that lead to BEPs. Although we do not have an answer, recent work of Nachbar (1997) on learning in games suggests that the set of prior beliefs for which learning is possible is quite small. A natural question for the reader to ask at this point (if not earlier) is why we focus on Bayesian learning. It is conceptually incredibly demanding, and it seems that - because of the need to tie priors together - this approach could never answer the learning question completely. We do not dispute this, and in fact we will argue that the Bayesian approach has already been pushed too far. Nonetheless, the Bayesian approach is surely the natural place to start. If a market participant is an expected utility maximizer then, as a consequence of this assumption, beliefs must be revised in light of new information according to Bayes's rule. Because Bayesian learning is a consequence of assumptions about preferences, it is frequently referred to as rational learning. We will follow this practice, but the reader should keep in mind that there is nothing necessarily irrational about ad hoc learning. To label non-Bayesian learning as irrational is to invest the Savage axioms that are at the foundation of expected utility maximization with normative content, which we (and we believe most economists) would reject. Game theory presents learning issues similar to the issues of expectation formation in economies with a sequence of incomplete markets and markets with differentially informed traders. In games with incomplete information, a (Bayes-Nash) equilibrium implies that, throughout the course of play, players will be learning. But many different structures of beliefs will be consistent with many, distinctly different, equilibria. Jordan (1991,1995) investigates the equilibrium behavior of infinitely repeated games. He demonstrates how the effects of learning force a relationship between limit beliefs in such a game and the equilibria of complete information versions of the game. Kalai and Lehrer (1993a) and Nyarko (1991, 1994b) ask whether players can learn their way to a Nash equilibrium when they do not necessarily start in a Bayes-Nash equilibrium. In different, but related, models they both provide positive answers. The issues that arise in this literature are essentially the same as those that arise in the microeconomic rational expectations literature. In this paper, our goal is not to survey the work on equilibrium under uncertainty or on the existence of rational expectations equilibrium, nor even to survey all the recent work on rational learning.1 Instead, our goal 1

Blume, Bray, and Easley (1982) provide a survey of learning in economies with differential information; Blume and Easley (1993) provide a partial survey of the recent work on learning in games; and Jordan (1992) provides an exposition of recent results on Bayesian learning in games and a non-Bayesian interpretation of some of these results.

64

Lawrence E. Blume and David Easley

is to provide an overview of the methods of analysis and results obtained and, most important, an assessment of the success of rational learning dynamics in tying down limit beliefs and limit behavior in gametheoretic and economic equilibrium models. We illustrate the features common to rational or Bayesian learning in single-agent, game-theoretic, and equilibrium frameworks. We show that rational learning is possible in each of these environments. The issue is not whether rational learning can occur, but rather what results it produces. The key to the success or failure of rational learning lies in identifying the proper parameter set. If we assume a natural complex parameterization of the choice environment, then all we know is the rational learner believes that his posteriors will converge somewhere with prior probability 1. Alternatively, if we as modelers assume the simple parameterization of the choice environment that is necessary to obtain positive results, then we are closing our models in the ad hoc fashion that rational learning was introduced to avoid. We do not believe that any further resolution of this issue is possible. Rational learning can indeed produce convergence of beliefs to rational expectations equilibria in competitive economies or to Nash equilibria in games. But in naturally complex environments, it can only do so if the modeler first imposes an equilibrium condition on the learning process. This equilibrium condition is weaker than rational expectations equilibrium or Bayes-Nash equilibrium, but it is nonetheless an equilibrium condition. We believe that a partial resolution of this conundrum is to pay more attention to how learning interacts with other dynamic forces. In the penultimate section of this paper we show that, in a simple economy, the forces of market selection can yield convergence to rational expectations equilibria even without every agent behaving as a rational learner. Similar evolutionary forces have been shown to be successful in producing Nash equilibrium or even refinements of Nash in repeated games. How successful such forces are in general, and how they interact with rational or nonrational learning, needs more study. We believe that this is a promising direction for future research. In the next section we discuss learning in the context of a single-agent decision problem. Along the way we introduce some of the tools that have proven useful in the analysis of learning dynamics. Section 3 discusses the role of learning in the analysis of repeated games, and Section 4 discusses learning in general equilibrium models. In Section 5 we discuss the robustness of learning in equilibrium models. Our conclusions about what we have learned from the learning literature are contained in Section 6.

Rational Expectations and Rational Learning 2

65

Learning Dynamics

The rational learning literature takes off from the analysis of Bayesian decision problems. Here we establish the basic results for the singleagent learning problem. The problem fundamental to the statistical literature is consistency. That is, will a decision maker ultimately learn the truth? We will introduce another problem which is important for equilibrium dynamics: the prediction problem. That is, does the prediction of the future path of the process, given its history through date t, converge to the correct conditional distribution as t grows? We will see that the relationship between consistency and the prediction problem is not as straightforward as it might seem. In this section we discuss the dynamics of Bayesian posterior belief revision and the problems of consistency and predication. We then describe a canonical decision problem, and discuss the problem of incomplete learning in some examples. 2.1

The Dynamics of Posterior Revision

Bayesian posterior revision works on a set of sample histories H = njli Ht (where Ht is the set of possible observations at time i), a set of parameters 0, and for each 6 e 0 a probability measure /ue on H. We assume that 0 and each Ht are Polish (complete, separable metric) spaces. We let 5* denote the product cr-field of subsets of H derived from the Borel a-fields on each Ht, and we assume that, for each event 5 e 5 , the map 6 -^jud(S) is Borel measurable. The Bayesian "learner" begins with a prior distribution v on 0. Corresponding to each prior v is the (unique) joint distribution 0V on 0 X H such that, for any set A X B with A a measurable subset of 0 and 4>*(A*B) = \AiAe(B)dv(e\ Just as v is the marginal distribution of (pv on 0, let /uv denote the marginal distribution of 0V on H. Posterior beliefs are just conditional distributions derived from 0V. Let HT = Hi X . .. X HT denote the set of possible observations through time T, and let a(hu . . . ,hT) be the a-field generated by the observation of hTeHT. The date T + l posterior distribution vr+1(- \hu • • ,h •T) is a regular version of